V. - A. - Marchenko - A. - Boutet - de - Monvel - H. - McKean - ( (BookFi) PDF

Mathematical Physics, Analysis and Geometry 2: 1–24, 1999.
1
© 1999 Kluwer Academic Publishers. Printed in the Netherlands.
Square Integrability and Uniqueness of the

Solutions of the Kadomtsev–Petviashvili-I Equation
LI-YENG SUNG
Department of Mathematics, University of South Carolina, Columbia, SC 29208, U.S.A.
(Received: 20 February 1998; in final form: 27 November 1998)

Abstract. We prove that the solution of the Cauchy problem for the Kadomtsev–Petviashvili-I
Equation obtained by the inverse spectral method belongs to the Sobolev space H k (R2 ) for k > 0,
under the assumption that the initial datum is a small Schwartz function. This solution is shown to be
the unique solution within a class of generalized solutions of the Kadomtsev–Petviashvili-I equation.
Mathematics Subject Classification (1991): 35Q53.
Key words: Kadomtsev–Petviashvili-I equation, inverse spectral method, Cauchy problem, unique-
ness of solutions.
1. Introduction
The Cauchy problem
(qt − 6qqx + qxxx )x = 3qyy , (1.1a)

q(x, y, 0) = q0 (x, y), (1.1b)
for the Kadomtsev–Petviashvili-I (KPI) equation (cf. [8]) is solved in [6] by the
inverse spectral method, under the assumption that q0 is a small Schwartz function.
It is shown in [6] that the solution q(x, y, t) obtained by the inverse spectral method
is a C ∞ classical solution of (1.1) for (x, y) ∈ R2 , t 6= 0, and t → q(·, t) ∈
C 1 ((−∞, 0) ∪ (0, ∞), C0 (R2 )).
We will show in this paper that t → q(·, t) ∈ C(R, H ∞ (R2 )) and q(x, y, t)
is the unique generalized solution for the forward (respectively backward) prob-
lem of (1.1) in the class C([0, ∞), H 3 (R2 )) ∩ C 1 ((0, ∞), L∞ (R2 )) (respectively
C((−∞, 0], H 3 (R2 )) ∩ C 1 ((−∞, 0), L∞ (R2 ))).
The inverse spectral method for KPI is studied formally in [10] and [5]. Rigor-
ous aspects of this method have been investigated in [14] and [20]. The version of
the inverse spectral method used in this paper is essentially that of [20]. However,
the results in [14] and [20] are obtained under the additional assumption that
Z
dx q0 (x, y) = 0, (1.2)
R
2 LI-YENG SUNG
which is also assumed in many of the papers that study (1.1) by PDE techniques
(cf. [4, 12, 13, 16, 17]). As a consequence of the nonphysical constraint (1.2), the
scattering data have decay in all directions, which greatly simplifies the analysis.
When (1.2) is not assumed, the analysis is much more subtle due to the lack of
decay of the scattering data in certain directions.
The implications of the constraint (1.2) are also studied in [1, 2] and [3] using
the inverse spectral method. But the fact that the solution obtained by the inverse
spectral method (without assuming (1.2)) is the unique solution of (1.1) in a general
class of solutions has not been rigorously established until now.
Since there is an isomorphism (cf. [9]) between solutions of the KP equation
and the Johnson equation (cf. [7]) in the case of rapidly decaying initial data, the
results of this paper can also be applied to the Johnson equation.
The rest of the paper is organized as follows. The inverse spectral method for
(1.1) is described in Section 2, where we give both the solution from the left and
the solution from the right. We also recall some relevant results from [6]. Section 3
contains the proof that the inverse spectral solution from the left and from the right
are identical. The integrability of the time-dependent Jost function is studied in
Section 4. We prove in Section 5 that t → q(·, t) ∈ C(R, H ∞ (R2 )), and establish
the uniqueness of solutions in Section 6.
For the convenience of the reader we collect here some notation frequently used
in this paper.
(a) S(Rn ) is the space of Schwartz functions in n real variables and S 0 (Rn ) is the
space of tempered distributions equipped with the weak∗ -topology.
(b) H k (R2 ), k = 0, 1, T
2, . . ., are the L2 based Sobolev spaces in two real variables,
∞
and H ∞ (R2 ) = k 2
k=0 H (R ) is equipped with the natural Fréchet space
m p
topology. Wp (R) is the L based Sobolev space in one real variable.
(c) C(X, Y ) is the space of continuous maps from the topological space X into the
topological space Y and Cb (X, Y ) is the space of bounded continuous maps
from the Banach space X into the Banach space Y . Cb (Rn ) is the space of
bounded continuous functions on Rn equipped with the sup-norm and C0 (Rn )
is the subspace of Cb (Rn ) whose members vanish at infinity.
(d) Let I be an open interval and Y be a topological vector space. C 1 (I, Y ) is the
space of continuously differentiable Y -valued functions on I .
(e) Let X be a measure space and Y be a Banach space. L1 (X, Y ) is the space of
Y -valued Bochner integrable functions on X.
(f) The Fourier transforms fˆ and f˜ are defined by
Z Z
ˆ
f (ξ, η) = dx dy e −i(xξ +yη) ˜
f (x, y) and f (ξ, y) = dx e−ixξ f (x, y).
R2 R
(g) The operators P± are defined by

Z
±1 f (l)
(P± f )(k) = dl .
2π i R l − (k ± 0i)
SOLUTIONS OF THE KADOMTSEV–PETVIASHVILI-I EQUATION 3
They always act on functions in the k variable in this paper.

(h) A . B means A 6 constant × B, where the constant is independent of the
variables in A and B.
2. The Inverse Spectral Method

We describe in this section the solution of (1.1) obtained by the inverse spectral
method. We consider the following time-dependent Schrödinger equation defined
by the initial datum q0 :
iφy − φxx + q0 φ = 0, (2.1)
and introduce the Jost function µ(x, y, k) by the relation
φ(x, y, k) = µ(x, y, k)e−i(xk−yk ) .

2
(2.2)
The equation for µ is
iµy − µxx + 2ikµx + q0 µ = 0. (2.3)
Taking the Fourier transform in the x variable, we can write (2.3) as

i
µ̃y (ξ, y, k) = i(ξ 2 − 2kξ )µ̃(ξ, y, k) + (q̃0 ∗ µ̃)(ξ, y, k), (2.4)
2π
where ∗ denotes the convolution in the ξ variable.
The Jost function µ+ (resp. µ− ) is defined for Im k > 0 (resp. Im k 6 0) by
Z y Z ∞
+ i 0 0
µ̃ (ξ, y, k) = 2π δ(ξ ) + E− (ξ ) dy − E+ (ξ ) dy ×
2π −∞ y
2 −2kξ )(y−y 0 )
× ei(ξ (q̃0 ∗ µ̃+ )(ξ, y 0 , k), (2.5)
Z y Z ∞
i
µ̃− (ξ, y, k) = 2π δ(ξ ) + E+ (ξ ) dy 0 − E− (ξ ) dy 0 ×
2π −∞ y
i(ξ 2 −2kξ )(y−y 0) − 0
×e (q̃0 ∗ µ̃ )(ξ, y , k), (2.6)
where δ is the Dirac function, and E± are defined by

1, ±ξ > 0,
E± (ξ ) = (2.7)
0, otherwise.
The left scattering data L± (k, l) (k, l ∈ R) are defined by

 Z
 i
dx dy e−i[x(k−l)+y(l −k )] q0 (x, y)µ− (x, y, k),
2 2
− k 6 l,
L (k, l) = 2π R2 (2.8)

0, otherwise,
4 LI-YENG SUNG
 Z
 i
dx dy e−i[x(k−l)+y(l −k )] q0 (x, y)µ+ (x, y, k),
2 2
k > l,
L+ (k, l) = 2π R2 (2.9)

0, otherwise.
Using L± we define the time-dependent Jost function from the left λ(x, y, t, k),
k ∈ R, by the following equation:
Z k Z ∞
− + + −
λ(x, y, t, k) = 1 + P dlL (k, l) + P dlL (k, l) ×
−∞ k
i[x(k−l)+y(l 2−k 2 )+4t (k 3 −l 3 )]
×e λ(x, y, t, l). (2.10)
We can now write down the solution of (1.1) from the left:
Z Z ∞ Z k
1 −
dlL (k, l) ei[x(k−l)+y(l −k )] ×
+ 2 2
q(x, y, t) = dk dlL (k, l) −
π R k −∞
3 −l 3 )
× e4it (k [i(k − l)λ(x, y, t, l) + λx (x, y, t, l)]. (2.11)
In summary, given a small q0 ∈ S(R2 ), the inverse spectral solution from the
left for (1.1) is obtained through the following steps: (I) Solve (2.5) and (2.6)
for µ± (x, y, k). (II) Define L± (k, l) by (2.8) and (2.9) using µ± (x, y, k). (III)
Solve (2.10) for λ(x, y, t, k). (IV) The solution q(x, y, t) is defined by (2.11) using
L± (k, l) and λ(x, y, t, k).
Note that we can also define the solution q(x, y, t) via the right scattering data:
 Z
 −i
dx dy e−i[x(k−l)+y(l −k )] q0 (x, y)µ− (x, y, k), k > l,
2 2
R −(k, l) = 2π R2 (2.12)

0, otherwise,
 Z
 −i
dx dy e−i[x(k−l)+y(l −k )] q0 (x, y)µ+ (x, y, k), k 6 l,
2 2
R +(k, l) = 2π R2 (2.13)

0, otherwise.
Let the time-dependent Jost function from the right ρ(x, y, t, k), k ∈ R, be defined
by the following equation:
Z ∞ Z k
− + + −
ρ(x, y, t, k) = 1 + P dlR (k, l) + P dlR (k, l) ×
k −∞
i[x(k−l)+y(l 2−k 2 )+4t (k 3 −l 3 )]
×e ρ(x, y, t, l). (2.14)
The solution of (1.1) from the right can then be expressed as
Z Z k Z ∞
1 −
dlR (k, l) ei[x(k−l)+y(l −k )] ×
+ 2 2
q(x, y, t) = dk dlR (k, l) −
π R −∞ k
3 −l 3 )
× e4it (k [i(k − l)ρ(x, y, t, l) + ρx (x, y, t, l)]. (2.15)
The relations among L± (k, l), R ± (k, l), λ(x, y, k) and ρ(x, y, k), and the equiv-
alence of (2.11) and (2.15) will be established in Section 3.
Finally we recall some relevant results from [6] which are obtained under the
smallness assumption
Z
dy dξ(1 + ξ 2 )|q̃0 (ξ, y)| 1. (2.16)
R2
(i) The integral equation (2.5) (resp. (2.6)) is uniquely solvable in the Banach
space Cb (Ry , L1 (Rξ ) ⊕ Cδ(ξ )) for Im k > 0 (resp. Im k 6 0).
(ii) The scattering data L− (k, l) and R + (k, l) (resp. L+ (k, l) and R − (k, l)) are
C ∞ for k 6 l (resp. k > l). Let F (k, l) be L± (k, l) or R ± (k, l). Then the
following estimates hold:
r α+β
∂ ∂ ∂
(k − l) (k − l )
m 2 2 n
+ F (k, l) . (1 + |k|)α+β
∂k ∂l ∂k∂l
for α, β, m, n > 0. (2.17)
(iii) In particular, we have
|F (k, l)| 6 Cq0 (1 + |k − l|)−2 , (2.18)
where Cq0 1 under the assumption (2.16), and the integral equations
(2.10) and (2.14) are uniquely solvable in L2 (Rk ) ⊕ C for (x, y, t) ∈ R3 .
(iv) The Jost functions µ± are related by
µ± (x, y, k) = µl (x, y, k) ∓
Z k
dl ei[x(k−l)+y(l −k )] L± (k, l)µl (x, y, l),
2 2
∓ (2.19)
∓∞
µ± (x, y, k) = µr (x, y, k) ∓
Z ±∞
dl ei[x(k−l)+y(l −k )] R ± (k, l)µr (x, y, l),
2 2
∓ (2.20)
k
where the left and right Jost functions µl and µr are the unique solutions of
µ̃l (ξ, y, k) = 2π δ(ξ ) +

Z y
i 0
dy 0 ei(ξ −2kξ )(y−y ) (q̃0 ∗ µ̃l )(ξ, y 0 , k),
2
+ (2.21)
2π −∞
µ̃r (ξ, y, k) = 2π δ(ξ ) −
Z ∞
i 0
dy 0 ei(ξ −2kξ )(y−y ) (q̃0 ∗ µ̃r )(ξ, y 0 , k).
2
− (2.22)
2π y
(v) Let # be l, r, +, or −. For αj > 0, we have the following estimates:
6 LI-YENG SUNG
∂ α1 +α2 +α3
(µ# (x, y, k) − 1) . (1 + |y|α3 ) for (x, y, k) ∈ R3 . (2.23)
∂x α1 ∂y α2 ∂k α3
(vi) Let q(x, y, t) be defined by either (2.11) or (2.15). Then we have
q(x, y, 0) = q0 (x, y). (2.24)
(vii) For t 6= 0, Equations (2.11) and (2.15) can be written as
Z Z ∞ Z k
∂ 1
q(x, y, t) = dk dlL− (k, l) − dlL+ (k, l) ×
∂x π R k −∞
!
i[x(k−l)+y(l 2 −k 2 )+4t (k 3 −l 3 )]
×e λ(x, y, t, l) , (2.25)
Z Z k Z ∞
∂ 1 − +
q(x, y, t) = dk dlR (k, l) − dlR (k, l) ×
∂x π R −∞ k
!
i[x(k−l)+y(l 2 −k 2 )+4t (k 3 −l 3 )]
×e ρ(x, y, t, l) , (2.26)
where the integrals exist as iterated integrals.

(viii) The function q(x, y, t) defined by (2.11) (resp. (2.15)) is C ∞ for t 6= 0, and
∂ m+n q
is continuous in R3 for m, n > 0. (2.27)
∂x m ∂y n
3. The Equivalence of the Left and Right Solutions

We first investigate the relations between the scattering data L± (k, l) and R ± (k, l).
Let the integral operators L± and R± be defined by
Z k
[L± f ](k) = ± dlL± (k, l)f (l) and
∓∞
Zk
[R± f ](k) = ∓ dlR ± (k, l)f (l). (3.1)
±∞
LEMMA 3.1. The operators I − L± and I − R± are bounded and invertible on

Lp (R) (1 6 p < ∞) and Cb (R).
Proof. Let F (k, l) = L± (k, l) or R ± (k, l). From (2.17) we have |F (k, l)| .
[1 + (k − l)2 ]−1 which implies the boundedness of L± and R± .
We will only show that L+ is invertible. Let f ∈ Lp (R) (1 6 p < ∞) or
Cb (R). From (2.17) we derive
Z Z
|f (l)|
dl|L+ (k, l)f (l)| . dl
R R (1 + |k 2 − l 2 |)2
Z
|f (l)|
. dl +
(1 + |k − l||k|)2
lk>0
Z
|f (l)|
+ dl . (3.2)
lk<0 (1 + |k + l||k|)2
Using (3.2) we obtain for any K > 0 that, for 1 6 p 6 ∞,
Z ∞ Z 0
|f (l)| |f (l)|
+
kL f kLp (k>K) . dl + dl
0 (1 + |k − l|K)2
−∞ (1 + |k + l|K) Lp (k>K)
2
Z ∞ Z 0
|f (l)| |f (l)|
. dl + dl
0 (1 + |k − l|K)2 (1 + |k + l|K) p
−∞
2
L (Rk )
1
. kf kLp (R) , (3.3)
K
and similarly
1
kL+ f kLp (k<−K) . kf kLp (R) . (3.4)
K
It follows from (3.3), (3.4), the smoothness of L+ (k, l) and the Fréchet–Kolmo-
gorov theorem (respectively the Ascoli–Arzelà theorem) (cf. [19]) that L+ is a
compact operator on Lp (R) for 1 6 p < ∞ (respectively Cb (R)).
We note that I − L+ has the same kernel on Lp (R) and Cb (R). Let f ∈ Cb (R)
belong to Ker(I − L+ ). We derive from (2.17) that for any K > 0
Z k
|f (l)|
kL+ f kL∞ (k<−K) . dl
−∞ (1 + |k 2 − l 2 |)2
Z k
|f (l)| 1
. dl . kf kL∞ (k<−K) . (3.5)
−∞ (1 + |k − l|K) 2 K
Since f (k) = (L+f )(k), it follows from (3.5) that f (k) = 0 for k sufficiently
negative, and then Gronwall’s inequality (cf. [18]) implies f (k) = 0 for all k ∈ R.
The invertibility of L+ follows from the Fredholm alternative (cf. [19]). 2
Let (L± )t and (R± )t be the transposes of L± and R± respectively, i.e.,

Z l
± t
[(L ) f ](l) = ∓ dkL± (k, l)f (k) and
±∞
Z l
[(R± )t f ](l) = ± dkR ± (k, l)f (k). (3.6)
∓∞
We have the following analog of Lemma 3.1.
LEMMA 3.2. The operators I − (L± )t and I − (R± )t are bounded and invertible
on Lp (R) (1 6 p < ∞) and Cb (R).
8 LI-YENG SUNG
We define
φ # (x, y, k) = e−i(xk−yk ) µ# (x, y, k),
2
where # = +, −, l, or r. (3.7)
Then (2.19) and (2.20) can be rewritten as
φ ± = (I − L± )φ l and φ ± = (I − R± )φ r . (3.8)
From (3.8) and Lemma 3.1 we derive the following equations in Cb (R):
φ r = (I − R+ )−1 (I − L+ )φ l and φ r = (I − R− )−1 (I − L− )φ l . (3.9)
For k ∈ R, let the functions ηl (x, y, k) be defined by the following analog
of (2.21):
Z y
i 0
dy 0 ei(ξ −2kξ )(y−y ) qe0 ∗ η̃l (ξ, y 0 , k), (3.10)
2
η̃l (ξ, y, k) = 2π δ(ξ ) +
2π −∞
which is uniquely solvable in Cb (Ry , L1 (Rξ )⊕Cδ(ξ )) under the assumption (2.16).
Of course ηl enjoys the same properties as µl .
Let γ l (x, y, k) be defined by
γ l (x, y, k) = e−i(xk−yk ) ηl (x, y, k),

2
(3.11)
then we have (cf. (2.1)–(2.4))
φyl = −iφxx
l
+ iq0 φ l and γyl = −iγxx
l
+ iq0 γ l . (3.12)
LEMMA 3.3. Let ζ(ξ, y, k) (resp. τ (ξ, y, k)) be the Fourier transform of
(µl (x, y, k)−1) (resp. the complex conjugate of (η(x, y, k) − 1)) in the x variable.
Then we have
Z Z
lim dξ sup |ζ(ξ, y, k)| = lim dξ sup |τ (ξ, y, k)| = 0. (3.13)
y→−∞ R k∈R y→−∞ R s∈R
Proof. It suffices to discuss the case of ζ(ξ, y, k). From (2.21) we obtain
Z y
0
dy 0 ei(ξ −2kξ )(y−y ) q̃0 (ξ, y 0 ) +
2
ζ(ξ, y, k) = i
−∞
Z y
i 0
dy 0 ei(ξ −2kξ )(y−y ) (q̃0 ∗ ζ )(ξ, y 0 , k).
2
+ (3.14)
2π −∞
It follows immediately from (3.14) that ζ(ξ, y, k) is continuous in all the vari-
ables and hence supy,k∈R |ζ(ξ, y, k)| is a measurable function in ξ . Moreover, we
find from the Neumann series solution of (3.14) that
∞
X
sup |ζ(ξ, y, k)| 6 ζj (ξ ), (3.15)
y,k∈R j =0
where
j Z j +1
1
kζj kL1 (R) 6 dξ dy|q̃0 (ξ, y)| . (3.16)
2π R2
Combining (2.16), (3.15) and (3.16) we have

Z
dξ sup |ζ(ξ, y, k)| < ∞. (3.17)
R y,k∈R
The first limit in (3.13) follows from (3.14) and (3.17). 2
LEMMA 3.4. Let f (k) ∈ S(R), then

Z Z
1
dxφ (x, y, k) dsγ l (x, y, s)f (s) = f (k) ∀y ∈ R.
l
(3.18)
2π R R
Proof. It follows from (2.23) and (3.12) that the iterated integral on the left-hand
side of (3.18) is independent of y. Using the notation in Lemma 3.3 we can also
rewrite the left-hand side of (3.18) by the Fourier inversion formula as
Z Z
1 l
dxφ (x, y, k) dsγ l (x, y, s)f (s)
2π R R
Z Z
1 −i(xk−yk 2 ) 2
= dx e (µ (x, y, k) − 1) ds ei(xs−ys ) (ηl (x, y, s) − 1)f (s)+
l
2π R R
Z Z
1 −i(xk−yk 2 ) 2
+ dx e (µ (x, y, k) − 1) ds ei(xs−ys ) f (s)+
l
2π R R
Z Z
1
dx e−i(xk−yk ) ds ei(xs−ys ) (ηl (x, y, s) − 1)f (s) + f (k)
2 2
+
2π R R
2 Z Z
eiyk
ds dξ ζ(k − s − ξ, y, k)τ (ξ, y, s)e−iys f (s)+
2
= 2
(2π ) R R
2 Z
eiyk
ds[ζ(k − s, y, k) + τ (k − s, y, s)]e−iys f (s) + f (k).
2
+ (3.19)
2π R
The lemma follows from (3.13) and (3.19). 2
LEMMA 3.5. The following identity holds as operators on L2 (R):

(I − R− )−1 (I − L− ) = (I − R+ )−1 (I − L+ ). (3.20)
Proof. From (3.9) we have (I − R− )−1 (I − L− )φ l = (I − R+ )−1 (I − L+ )φ l . Let

g ∈ L1 (R). Then we have
Z Z
dk φ (x, y, k)G− (k) =
l
dk φ l (x, y, k)G+ (k), (3.21)
R R
where
10 LI-YENG SUNG
G± = [I − (L± )t ][I − (R± )t ]−1 g. (3.22)
Since G± ∈ L1 (R) by Lemma 3.2 and φ l (x, y, k) is bounded and continuous in

all the variables, we find by (3.18), (3.21) and Fubini’s theorem that
Z Z
dkf (k)G− (k) = dkf (k)G+ (k) ∀f ∈ S(R). (3.23)
R R
It follows from (3.23) that

Z
dk[(I − R− )−1 (I − L− )f ](k)g(k)
R Z
= dk[(I − R+ )−1 (I − L+ )f ](k)g(k). (3.24)
R
Since g ∈ L1 (R) is arbitrary and S(R) is dense in L2 (R), the identity (3.20)
follows from (3.24). 2
We now show that (2.11) and (2.15) define the same function q(x, y, t). We can
rewrite (2.10) and (2.14) as
−
P E(x,y,t )(I − L+ )E−1 + − −1
(x,y,t ) + P E(x,y,t )(I − L )E(x,y,t ) λ = 1, (3.25)
−
P E(x,y,t )(I − R+ )E−1 + − −1
(x,y,t ) + P E(x,y,t )(I − R )E(x,y,t ) ρ = 1, (3.26)
where [E(x,y,t )f ](k) = ei(xk−yk +4t k ) f (k).

2 3
Combining (3.20), (3.25) and (3.26) we obtain the following lemma.
LEMMA 3.6. The functions λ(x, y, t, k) and ρ(x, y, t, k) are related by

ρ = E(x,y,t )(I − R+ )−1 (I − L+ )E−1
(x,y,t )λ
= E(x,y,t )(I − R− )−1 (I − L− )E−1

(x,y,t )λ. (3.27)
PROPOSITION 3.7. The formulas (2.11) and (2.15) define the same function
q(x, y, t).
Proof. The case where t = 0 follows from (2.24). For t 6= 0, in view of (2.25)
and (2.26), we can rewrite (2.11) and (2.15) as
q(x, y, t)
Z
1 ∂ + − −1

= dk E(x,y,t )((I − L ) − (I − L ))E(x,y,t ) λ(x, y, t, l) , (3.28)
π ∂x R
q(x, y, t)
Z
1 ∂ + − −1

= dk E(x,y,t )((I − R ) − (I − R ))E(x,y,t ) ρ(x, y, t, l) . (3.29)
π ∂x R
The proposition follows from (3.27)–(3.29). 2

4. The Integrability of λ(x, y, t, k) and ρ(x, y, t, k) in the (x, k) Variables

We study the integrability of functions related to the time-dependent Jost functions
in this section. The results of this section are crucial for proving in the next section
that t → q(·, t) ∈ C(R, H ∞ (R2 )). We will concentrate on λ(x, y, t, k).
Let L(y,t ) be the operator defined by
Z k
(L(y,t )f )(x, k) = P− dlL+ (k, l) +
−∞
Z ∞
+
dlL (k, l) ei[x(k−l)+y(l −k )+4t (k −l )] f (x, l). (4.1)
− 2 2 3 3
+P
k
LEMMA 4.1. For each (y, t), L(y,t ) is a contraction on L2 (A × Rk ) for any A ⊆
Rx .
Proof. We have, by (2.18),
Z k

∓
dlL (k, l)e i[x(k−l)+y(l 2−k 2 )+4t (k 3 −l 3 )]
f (x, l)
2
±∞ L (A×Rk )
Z

6 Cq 0 −2
dl(1 + |k − l|) kf (·, l)kL2 (A)
. (4.2)
R L2 (Rk )
The lemma follows from (4.2) and the smallness of Cq0 . 2
Next we have a technical lemma for functions in the Sobolev spaces Wpm (R).
LEMMA 4.2. Let 1 < p < ∞ and f (k) ∈ Wpm (R) for m > 0. Then
± ∓isk
P (e f (k)) m . (1 + s)−j kf k m+j+1 for s > 0 and j, m > 0. (4.3)
W (R) p Wp (R)
Proof. We will only prove the estimate for g = P+(e−isk f (k)). Also, it suffices
to prove (4.3) for j = 0 since
sP+ (e−isk f (k)) = −iP+ (e−isk f 0 (k)) + i[P+ (e−isk f (k))]0 .
For m = 0 or 0 6 s < 1, this follows immediately from the properties of the

Hilbert transform (cf. [15]). So we assume m, s > 1.
Since s > 0, a simple contour integration gives
Z ∞
1 e−is(l−k) 1
p.v. dl =− . (4.4)
2π i −∞ l−k 2
On the other hand we obtain from the Plemelj formula (cf. [11])
Z ∞
1 f (l) 1
g(k) = p.v. dl e−isl + e−isk f (k). (4.5)
2π i −∞ l−k 2
12 LI-YENG SUNG
It follows from (4.4) and (4.5) that

Z
1 f (l) − f (k)
g(k) = (R) dl e−isl , (4.6)
2π i R l−k
R
where (R) dl denotes the improper Riemann integral.
A simple calculation using integration by parts shows that
Z
∂m e−isl
(R) dl
∂k m |l|>L (l − k)
Z −isl
∂m e
= (R) dl m for m > 0, |k| < L. (4.7)
|l|>L ∂k (l − k)
By the partition R = {l : |l| 6 L} ∪ {l : |l| > L} and (4.7) we see that
Z m
1 −isl ∂ f (l) − f (k)
g (k) =
(m)
(R) dl e
2π i ∂k m l−k
"Z R
m
1 −isl ∂ f (l) − f (k)
= dl e +
2π i |l−k|<1 ∂k m l−k
Z
m!f (l)
+ dl e−isl −
|l−k|>1 (l − k)m+1
Z #
X m
m!f (j )
(k)
− (R) dl e−isl
|l−k|>1 j =0
j !(l − k)m+1−j
= g1 (k) + g2 (k) + g3 (k). (4.8)
R
We have trivially kg2 kLp . kf kLp since m > 1. The integrals (R) |l−k|>1 dl
e−isl
(l−k)m+1−j
,0 6 j 6 m, are uniformly bounded for s > 1 and k ∈ R, which implies
that kg3 k . kf kWpm .
Lp
We find by Taylor’s formula that
Z
∂ m f (l) − f (k)
dl e−isl m
|l−k|<1 ∂k l−k
Z 1 Z
e−is(l+k) l
= dl m+1 dt (l − t)m f (m+1) (t + k). (4.9)
−1 l m! 0
Hence we also have kg1 kLp . kf kWpm+1 . 2
In the following two lemmas the function F (k, l) is piecewise C ∞ on the half
planes {(k, l) : k > l} and {(k, l) : k 6 l}, and it satisfies the estimates in (2.17).
Rk
LEMMA 4.3. Let H (x, y, t, k) be either P− −∞ dlF (k, l)ei[x(k−l)+y(l −k )+4t (k −l )]
2 2 3 3
R ∞
or P+ k dlF (k, l)ei[x(k−l)+y(l −k )+4t (k −l )] , and H̃ (ξ, y, t, k) be the Fourier trans-
2 2 3 3
form of H (x, y, t, k) in the x variable. Then, for 1 < p < ∞ and m > 0,
Z
1 + |t|m+2
dξ kH̃ (ξ, y, t, k)kWpm (Rk ) . for t ∈ R, y 6 0. (4.10)
R 1 + |y|1−(1/p)
Moreover, if B ⊂ R2 is bounded and ε > 0, then

{H̃ (ξ, y, t, k) : ε < |ξ | < ε −1 , (y, t) ∈ B}
is a precompact subset of Wpm (Rk ). (4.11)
Proof. We will treat only the case where H is defined by the first formula. By a
change of variables we can write
Z ∞
−
H (x, y, t, k) = P dl eixl e−ik(2yl) G(l, y, t, k), (4.12)
0
where
2 +4t l 3 ) 2 −l 2 k)
G(l, y, t, k) = ei(yl F (k, k − l)e12it (lk . (4.13)
The estimates in (2.17) and the piecewise smoothness of F imply that, for each
(y, t), the map l → e−ik(2yl)G(l, y, t, k) belongs to C((0, ∞), L2 (Rk ))∩L1 ((0, ∞),
L2 (Rk )). Therefore the bounded operator P− can be moved inside the integral in
(4.12) and we obtain from the Fourier inversion formula that

H̃ (ξ, y, t, k) = 2π E+ (ξ )P− e−ik(2yξ ) G(ξ, y, t, k) , (4.14)
where E+ (·) is defined in (2.7).
From (2.17) we also obtain the following estimate:
` `
∂
G(l, y, t, k) . 1 + |t| min 1, 1 for ` > 0. (4.15)
∂k ` 1 + |kl| 1 + |l|
Combining (4.3) (with j = 0, 1) and (4.15) we find, for ξ > 0 and y 6 0,
− −ik(2yξ )
P e G(ξ, y, t, k) W m
p
m+1
1 + |t| 1 + |t| 1
. min , . (4.16)
|ξ |1/p 1 + |yξ | 1 + |ξ |
The estimate (4.10) follows immediately from (4.14) and (4.16).
The estimate (4.15) and the Lebesgue dominated convergence theorem imply
that the map (ξ, y, t) → e−ik(2yξ ) G(ξ, y, t, k) is continuous from the set {(ξ, y, t) :
ξ 6= 0} into Wpm (R). The precompactness statement follows immediately from the
boundedness of P− on Wpm (R). 2
LEMMA 4.4. Let V (x, y, t, k) be either

Z k
−
dl(k − l)F (k, l)ei[x(k−l)+y(l −k )+4t (k −l )]
2 2 3 3
P
−∞
14 LI-YENG SUNG
or
Z ∞
P+
2−k 2 )+4t (k 3 −l 3 )]
dl(k − l)F (k, l)ei[x(k−l)+y(l .
k
Then the following estimate holds:

1 + |t|3
kV (x, y, t, k)kL2 (R2x,k ) . for t ∈ R, y 6 0. (4.17)
1 + |y|
Moreover, if B is a bounded subset of R2 , we have
lim kV (x, y, t, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B, (4.18)

r→∞
where
r = {(x, k) ∈ R2 : |x| > r}. (4.19)

Proof. It suffices to discuss the case where V is defined by the first formula. We
have
Z ∞

V (x, y, t, k) = dl eixl P− e−ik(2yl) lG(l, y, t, k) , (4.20)
0
where G is defined in (4.13).

Applying (4.3) (with j = 0, 2) we obtain the following analog of (4.16) for
l > 0 and y 6 0:
− −ik(2yl)
P e lG(l, y, t, k) L2 (Rk )

1 + t2 1
. |l|1/2 (1 + |t|) min , . (4.21)
(1 + |2yl|)2 1 + l 2
The estimate (4.17) follows from (4.20), (4.21) and the Plancherel theorem.
From (4.15) (with ` = 0) we see that the map (y, t) → e−ik(2yl) lG(l, y, t, k)
is continuous from R2y,t to L2 (R2l.k ). It follows from (4.20), the boundedness of
P− on L2 (R) and the Plancherel theorem that {V− (x, y, t, k) : (y, t) ∈ B} is
also a precompact subset of L2 (R2x,k ), which then implies (4.18) by the Fréchet–
Kolmogorov theorem. 2
LEMMA 4.5. Let H (x, y, t, k) be as in Lemma 4.3. Then

1 + t4
kL(y,t )H kL2 (R2x,k ) . for t ∈ R, y 6 0. (4.22)
1 + |y|
Moreover, if B is a bounded subset of R2 and r is defined by (4.19), we have
lim k(L(y,t )H )(x, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B. (4.23)

r→∞
Proof. We have
Z k Z ∞
[L(y,t )H ](x, k) = P− +
dlL (k, l) + P + −
dlL (k, l) ×
−∞ k
i[x(k−l)+y(l 2 −k 2 )+4t (k 3 −l 3 )]
×e H (x, y, t, l)
= g− (x, y, t, k) + g+ (x, y, t, k), (4.24)
where
g∓ (x, y, t, k)
Z Z
1
= dξ eixξ
dl eixl E± (l)P∓[K± (l, y, t, k)H̃ (ξ, y, t, k − l)], (4.25)
2π R R
and
2 +4t l 3 )
e−ik(2yl) L± (k, k − l)e12it (lk
2 −l 2 k)
K± (l, y, t, k) = ei(yl . (4.26)
We note that (2.17) and (4.10) together imply (ξ, l) → K± (l, y, t, k)H̃ (ξ, y, t,
k−l) ∈ L1 (R2ξ,l , L2 (Rk )) for fixed (y, t). Therefore by integrating against arbitrary
L2 (Rk ) functions we obtain (4.25) as an identity on C(Rx , L2 (Rk )) for each fixed
(y, t).
From (2.17) and (4.3) we obtain the following estimate for y 6 0:
∓
P [K± (l, y, t, k)H̃ (ξ, y, t, k − l)] 2
L (Rk )

. (1 + t 2 )kH̃ (ξ, y, t, ·)kL2 (R) + (1 + |t|)kH̃ (ξ, y, t, ·)kH 1 (R) +

1 1
+ kH̃ (ξ, y, t, ·)kH 2 (R) min , . (4.27)
1 + |yl| 1 + |l|
Combining (4.25), (4.27), (4.10) (with p = 2) and the Plancherel theorem we find
Z
kg∓ kL2 (Rx,k ) . dξ kP∓ [K± (l, y, t, k)H̃ (ξ, y, t, k − l)]kL2 (R2 )
k,l
R
1+t 4
. for y 6 0. (4.28)
1 + |y|
The estimate (4.22) follows from (4.24) and (4.28).
Let ε be an arbitrary positive number. From (4.25) we have
g∓ (x, y, t, k)
Z Z
1
= dξ + dξ eixξ ×
2π |ξ |<ε |ξ |>ε −1
Z
× dl e E± (l)P∓[K± (l, y, t, k)H̃ (ξ, y, t, k − l)]+
ixl
R
Z Z
1
+ dξ e ixξ
dl eixl E± (l)P∓ [K± (l, y, t, k)H̃ (ξ, y, t, k − l)]
2π ε<|ξ |<ε−1 R
= I1 (x, y, t, k) + I2 (x, y, t, k). (4.29)
16 LI-YENG SUNG
It follows from (4.14), (4.15) and (4.27) that for bounded t we can make
kI1 (x, y, t, k)kL2 (Rx,k ) arbitrarily small by choosing ε small enough.
On the other hand, since |K± (l, y, t, k)| . (1 + |l|)−1 by (2.17), the map
(y, t) → K± (l, y, t, k)f (k) ∈ C(R2 , L2 (R2k,l )) for any f ∈ L2 (R). This together
with (4.11) show that P± [K(l, y, t, k)H̃ (ξ, y, t, k − l)] forms a precompact subset
of L2 (R2l,k ) for ε < |ξ | < ε −1 and (y, t) ∈ B.
R
The Plancherel theorem then implies R dl eixl E± (l)P∓[K(l, y, t, k)H̃ (ξ, y, t,
k − l)] forms a precompact subset of L2 (Rx,k ) for ε < |ξ | < ε −1 and (y, t) ∈ B.
Hence, we have
Z

dl e E± (l)P [K(l, y, t, k)H̃ (ξ, y, t, k − l)]
ixl ∓
→0 (4.30)

R L2 (r )
uniformly for ε < |ξ | < ε −1 and (y, t) ∈ B as r → ∞.

The estimate (4.30) implies that the L2 (r ) norm of I2 (x, y, t, k) tends to zero
uniformly for (y, t) ∈ B as r → ∞, and (4.23) follows. 2
LEMMA 4.6. The following estimate holds:

1 + t4
kλ(x, y, t, k) − 1 − (L(y,t )1)(x, k)kL2 (R2x,k ) . (4.31)
1 + |y|
for t ∈ R and y 6 0. Moreover, if B is a bounded subset of R2 and r is defined
by (4.19), we have
lim kλ(x, y, t, k) − 1 − (L(y,t )1)(x, k)kL2 (r ) = 0

r→∞
uniformly for (y, t) ∈ B. (4.32)
Proof. Let χy,t (x, k) = (1 + |y|)[λ(x, y, t, k) − 1 − (L(y,t )1)(x, k)]. Then χ(y,t )
satisfies
χ(y,t ) = (1 + |y|)L2(y,t )1 + L(y,t )χ(y,t ). (4.33)
Note that L(y,t )1 has the same properties as H in Lemma 4.3. The estimate (4.31)
then follows from (4.22) and (4.33).
Let ωy,t (x, k) = λ(x, y, t, k) − 1 − (L(y,t )1)(x, k), then we have
ω(y,t ) = L2(y,t )1 + L(y,t )ω(y,t ). (4.34)
The limit (4.32) follows from Lemma 4.1, Lemma 4.3, (4.23) and (4.34). 2
LEMMA 4.7. The following estimates hold:

j
∂ Pj (|t|)

∂x j [λ(x, y, t, k)] 2 2 6 1 + |y| for t ∈ R, y 6 0, j > 1, (4.35)
L (R ) x,k
where Pj is a polynomial. Moreover, if B is a bounded subset of R2 and r is

defined by (4.19), we have
j
∂
lim
[λ(x, y, t, k)]
2 = 0 uniformly for (y, t) ∈ B. (4.36)
r→∞ ∂x j
L (r )
Proof. Differentiating (2.10) we find
λx = h− + h+ + L(y,t )λx , (4.37)

where
Z k
h∓ (x, y, t, k) = ±P∓
2 −k 2 )+4t (k 3 −l 3 )]
dlei[x(k−l)+y(l ×
∓∞
× i(k − l)L± (k, l)λ(x, y, t, l). (4.38)
We can write h− (x, y, t, k) as

h− (x, y, t, k) = h1 (x, y, t, k) + h2 (x, y, t, k) + h3 (x, y, t, k), (4.39)
where
Z k
h1 (x, y, t, k) = P−
2−k 2 )+4t (k 3 −l 3 )]
dl ei[x(k−l)+y(l i(k − l)L+ (k, l), (4.40)
−∞
Z k
h2 (x, y, t, k) = P−
2−k 2 )+4t (k 3 −l 3 )]
dl ei[x(k−l)+y(l ×
−∞
× i(k − l)L+ (k, l)(L(y,t )1)(x, l), (4.41)
Z k
h3 (x, y, t, k) = P− dl ei[x(k−l)+y(l −k )+4t (k −l )] i(k − l)L+ (k, l) ×
2 2 3 3
−∞
× [λ(x, y, t, l) − 1 − (L(y,t )1)(x, l)]. (4.42)
We can estimate h1 (x, y, t, k) by Lemma 4.4, h2 (x, y, t, k) by Lemma 4.5 (with

L± (k, l) replaced by i(k − l)L± (k, l) in (4.24)), and h3 (x, y, t, k) by Lemma 4.6.
The results are, for j = 1, 2, 3,

1 + t4
khj (x, y, t, k)kL2 (R2x,k ) . for y 6 0, t ∈ R, (4.43)
1 + |y|
lim khj (x, y, t, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B. (4.44)
r→∞
For the case j = 1, the lemma follows from Lemma 4.1, (4.37), (4.39), (4.43),
(4.44) and the corresponding results for h+ (x, y, t, k). The higher order cases are
established by similar techniques and mathematical induction. 2
The following proposition is obtained by using Lemmas 4.1–4.5, mathematical

induction and the techniques in the proofs of Lemmas 4.6 and 4.7.
18 LI-YENG SUNG
PROPOSITION 4.8. Lemma 4.7 (resp. Lemma 4.6) is valid if λ(x, y, t, k) is re-
∂`
placed by ∂y ` λ(x, y, t, k) (resp. λ(x, y, t, k) − 1 − (L(y,t )1)(x, k) is replaced by
∂ j+`
∂x j ∂y `
[λ(x, y, t, k) − 1 − (L(y,t )1)(x, k)]) for ` > 0 (resp. j, ` > 0).
REMARK 4.9. We have so far obtained results concerning λ(x, y, t). But of course
similar results (for y > 0) also hold for ρ(x, y, k, t). This is the reason why we
need both (2.11) and (2.15) to represent the solution.
5. The Square Integrability of q(x, y, t) in the (x, y) Variables

We first establish a useful estimate. Let F (k, l) satisfy the estimates in (2.17) and
f ∈ L2 (R). Then we have
Z Z
1 |f (l)|
dk dl|F (k, l)f (l)| . dk dl
R2 R2 1 + |k − l| 1 + |k 2 − l 2 |
Z
1 |f ((k + l)/2)|
. dk dl
R2 1 + |k| 1 + |kl|
. kf kL2 (R) . (5.1)
We can now prove one of the main results of this paper.
THEOREM 5.1. Let q0 be a Schwartz function which satisfies the smallness as-
sumption (2.16) and q(x, y, t) be the solution of (1.1) obtained by the inverse
spectral method. Then for each t, q(·, t) ∈ H j (R2 ) for j > 0, and the map
t → q(·, t) is continuous from R into H j (R2 ).
Proof. Throughout this proof Qj is a polynomial in one real variable. From
(2.11) we can write
q(x, y, t) = q1,+ (x, y, t) + q1,− (x, y, t) + q2,+ (x, y, t) + q2,− (x, y, t),(5.2)
where
Z Z ±∞
i
dl ei[x(k−l)+y(l −k )+4t (k −l )] ×
2 2 3 3
q1,∓ (x, y, t) = dk
π R k
× (k − l)L∓ (k, l)λ(x, y, t, l), (5.3)
q2,∓ (x, y, t)
Z Z ±∞
1
dl ei[x(k−l)+y(l −k )+4t (k −l )] L∓ (k, l)λx (x, y, t, l). (5.4)
2 2 3 3
= dk
π R k
It follows from (5.1), (4.35) and (5.4) that

kq2,± (x, y, t)kL2 (Rx ) 6 Q1 (|t|)(1 + |y|)−1 for t ∈ R, y 6 0, (5.5)
and hence
kq2,± (x, y, t)kL2 ({(x,y):y 60}) 6 Q2 (|t|) for t ∈ R. (5.6)
In order to estimate q1,− (x, y, t), we break it into three terms:

X
3
q1,− (x, y, t) = vj (x, y, t), (5.7)
j =1
where
Z Z ∞
i 2−k 2 )+4t (k 3 −l 3 )]
v1 (x, y, t) = dk dl ei[x(k−l)+y(l (k − l)L− (k, l), (5.8)
π R k
Z Z ∞
i 2−k 2 )+4t (k 3 −l 3 )]
v2 (x, y, t) = dk dl ei[x(k−l)+y(l ×
π R k
× (k − l)L− (k, l)(L(y,t )1)(x, l), (5.9)

Z Z ∞
i
dl ei[x(k−l)+y(l −k )+4t (k −l )] (k − l)L− (k, l) ×
2 2 3 3
v3 (x, y, t) = dk
π R k
× [λ(x, y, t, l) − 1 − (L(y,t )1)(x, l)]. (5.10)
As in the case of q2,± , by applying (5.1) and (4.31) to (5.10) we obtain
kv3 (x, y, t)kL2 (Rx ) 6 Q3 (|t|)(1 + |y|)−1 for t ∈ R, y 6 0, (5.11)
and hence
kv3 (x, y, t)kL2 ({(x,y):y 60}) 6 Q4 (|t|) for t ∈ R. (5.12)
From (2.17), (5.8) and the change of variables ξ = k − l and η = l 2 − k 2 we
find
Z
dξ dηG(ξ, η)ei(ξ +3η /ξ )t ei(xξ +yη) ,
3 2
v1 (x, y, t) = (5.13)
R2
where kGkL2 (R2 ) < ∞. It then follows from (5.13) and the Plancherel theorem that
kv1 (x, y, t)kL2 (R2x,y ) . 1 for t ∈ R. (5.14)
Next we investigate the square integrability of v2 (x, y, t). Let H (k, l) = i(k −
l)L−(k, l) and M(ξ, y, t, k) be the Fourier transform of (L(y,t )1)(x, k) in the x
variable. Note that the results for H̃ (ξ, y, t, k) in Lemma 4.3 are also valid for
M(ξ, y, t, k).
We can rewrite v2 (x, y, t) as
Z Z
1
dl dk ei(xl+yl +4t l ) e−i2kly e12it (lk −l k) ×
2 3 2 2
v2 (x, y, t) = 2
dξ eixξ
2π R R2
× E− (l)H (k, k − l)M(ξ, y, t, k − l), (5.15)
From (2.17) we have
20 LI-YENG SUNG
Z Z 3/2
dl dk|H (k, k − l)| 4/3
< ∞. (5.16)
R R
By the Plancherel theorem, Lemma 4.3, (5.15), (5.16) and Hölder’s inequality we
have
Z Z

kv2 (x, y, t)kL2 (Rx ) . dξ
dk|H (k, k − l)M(ξ, y, t, k − l)|
R R L2 (Rl )
Z Z 3/2 1/2
. dl dk|H (k, k − l)|4/3 ×
R R
Z
× dξ kM(ξ, y, t, ·)kL4 (R)
R
6 Q5 (t)(1 + |y|)−3/4 for t ∈ R, y 6 0. (5.17)
kv2 (x, y, t)kL2 ({(x,y):y 60}) 6 Q6 (|t|) for t ∈ R. (5.18)
Moreover, by splitting the integral in (5.15) over the sets {ξ : |ξ | < ε −1 or
ε < |ξ |} and {ξ : ε < |ξ | < ε −1 }, we obtain from Lemma 4.3 (cf. the arguments in
the proof of Lemma 4.5) that, for any bounded subset B of R2 ,
lim kv2 (x, y, t)kL2 (|x|>r) = 0 uniformly for (y, t) ∈ B. (5.19)
r→∞
It follows from (5.7), (5.12), (5.14), (5.18) and their analogs for q1,+ that
kq1,± (x, y, t)kL2 ({(x,y):y 60}) 6 Q7 (|t|) for t ∈ R. (5.20)
The estimates (5.6) and (5.20) together with (5.2) show that
kq(x, y, t)kL2 ({(x,y):y 60}) 6 Q8 (|t|) for t ∈ R. (5.21)
Similarly, using Proposition 3.7 and Remark 4.9, we find
kq(x, y, t)kL2 ({(x,y):y >0}) 6 Q9 (|t|) for t ∈ R. (5.22)
From (5.5), (5.11), (5.13) and (5.17) we see that the L2 ({(x, y) : y 6 0}) norm
of q(x, y, t) becomes arbitrarily small for |y| large and t bounded. So to prove
the continuous dependence of q(x, y, t) in t with respect to L2 ({(x, y) : y 6 0}),
it suffices to look at the case where (y, t) ∈ B and B is a bounded subset of
{(y, t) : y 6 0}. Then (4.32), (4.36), (5.13) and (5.19) further show that we may
also assume x be bounded. The continuity of q(x, y, t) in t with respect to the
L2 ({(x, y) : y 6 0}) norm therefore follows from (2.27). In view of Remark 4.9
and Proposition 3.7 this is also true in the L2 (R2x,y ) norm.
We have proved the case where j = 0. The other cases are established by similar
techniques using Lemmas 4.3–4.5 and Proposition 4.8. 2
6. The Uniqueness of Solutions

The following lemma whose proof can be found in [6] enables us to write (1.1) in
a different form.
LEMMA 6.1. Suppose f, g ∈ L2 (R2 ) and h ∈ L∞ (R2 ) satisfy fy = (g + h)x in
the sense of distributions and the weak derivative fy ∈ L2 (R2 ), then there exists
8 ∈ S 0 (R2 ) such that 8x = f and 8y = g + h in the sense of distributions.
Moreover, if f , g and h depend continuously on some parameters, then 8 also
depends continuously on the same parameters.
Let q(x, y, t) ∈ C([0, ∞), H 3 (R2 )) ∩ C 1 ((0, ∞), L∞ (R2 )) be a generalized
solution of (1.1) for t > 0. By applying Lemma 6.1 to f = 3qy , g = qxxx − 6qqx
and h = qt , we obtain
qt − 6qqx + qxxx = 3ry , (6.1)
rx = qy , (6.2)
in the sense of distributions for t > 0, where t → r(·, t) belongs to C((0, ∞),
S 0 (R2 )).
REMARK 6.2. For each t, the distribution r(·, t) is only unique up to a constant.
The uniqueness of the generalized solution for the forward problem of (1.1)
in the class C([0, ∞), H 3 (R2 )) ∩ C 1 ((0, ∞), L∞ (R2 )) therefore follows from the
next proposition.
PROPOSITION 6.3. The Cauchy problem (1.1) has at most one generalized solu-
tion q in the sense of (6.1)–(6.2) for t > 0 such that

t → q(·, t) ∈ C [0, ∞), H 3 (R2 ) ∩ C 1 (0, ∞), S 0 (R2 ) , (6.3)
0

t → r(·, t) ∈ C (0, ∞), S (R ) . 2
(6.4)
Proof. The following arguments generalize those in [17]. We denote the L2 (R2 )
inner product and norm by (·, ·) and k · k respectively.
Let ε be an arbitrary positive number, βε (ξ, η) be a C ∞ function such that (i)
0 6 βε (ξ, η) 6 1, (ii) βε (ξ, η) = 1 on the set Eε = {(ξ, η) ∈ R2 : ε < |ξ | < ε −1
and |η| < ε −1 }, and (iii) βε (ξ, η) = 0 on R2 \ Eε/2 . Let αε ∈ S(R2 ) be the function
such that α̂ε = βε .
Assume that q1 (x, y, t) and q2 (x, y, t) are two such solutions of (1.1). We define
qj,ε , (qj2 )ε and rj,ε to be the convolutions of qj , qj2 and rj with αε . Then for j = 1, 2
we obtain from (6.1)–(6.4)
d
(qj,ε , φ) + 3 (qj2 )ε , φx − (qj,ε , φxxx ) = −3(rj,ε , φy )
dt
∀φ(x, y) ∈ S(R2 ), (6.5)
(rj,ε , φx ) = (qj,ε , φy )
∀φ(x, y) ∈ S(R2 ). (6.6)
22 LI-YENG SUNG
From (6.6) we obtain
r̂j,ε (ξ, η, t) = (η/ξ )βε (ξ, η)q̂(ξ, η, t), (6.7)
which together with (6.3) imply that

t → rj,ε (·, t) ∈ C (0, ∞), H ∞ (R2 ) . (6.8)
It then follows from (6.3), (6.5) and (6.8) that

t → qj,ε (·, t) ∈ C 1 (0, ∞), H ∞ (R2 ) ∩ C [0, ∞), H ∞ (R2 ) . (6.9)
Let 1ε (x, y, t) = q1,ε (x, y, t) − q2,ε (x, y, t), ωε (x, y, t) = (q12 )ε (x, y, t) −
(q2 )ε (x, y, t), and Rε (x, y, t) = r1,ε (x, y, t) − r2,ε (x, y, t). From (6.5), (6.6), (6.8)
2
and (6.9) we have
((1ε )t , 1ε ) = −3(ωε , (1ε )x ) + (1ε , (1ε )xxx ) − 3(Rε , (Rε )x ). (6.10)
Taking the real part of (6.10) we find dtd (1ε , 1ε ) = −3Re(ωε , (1ε )x ), which
together with (6.9) imply that
Z t
k1ε (·, t)k2 = −3 dt 0 Re(ωε (·, t 0 ), (1ε )x (·, t 0 )). (6.11)
0
Letting ε → 0 we obtain from (6.11)

Z t
k1(·, t)k = −3
2
dt 0 Re(σ (·, t 0 )1(·, t 0 ), 1x (·, t 0 )), (6.12)
0
where 1(x, y, t) = q1 (x, y, t)−q2 (x, y, t) and σ (x, y, t) = q1 (x, y, t)+q2 (x, y, t).
Let T > 0 be arbitrary. It follows from (6.3), (6.12) and the Sobolev inequality
(cf. [18]) that
Z t

k1(·, t)k 6 CT
2
dt 0 k1(·, t 0 )k2 + k1x (·, t 0 )k2 for 0 6 t 6 T . (6.13)
0
A similar argument yields

Z t
0

k1x (·, t)k 6 CT
2
dt 0 k1(·, t 0 )k + k1x (·, t 0 )k k1xx (·, t)k
0
for 0 6 t 6 T . (6.14)
From (6.5) and (6.6) we also find, for R = r1 − r2 ,
(1xx , 1xx ) + 3(Rx , Rx ) = 3Re(σ 1, 1xx ) for t 6= 0. (6.15)
It follows from (6.3), (6.15) and the Sobolev inequality that
k1xx k . k1k for t 6= 0. (6.16)

We deduce from (6.13), (6.14), (6.16) and Gronwall’s inequality that
k1k = k1x k = 0 for 0 6 t 6 T . (6.17)

2
It is shown in [6] that the function q(x, y, t) obtained by the inverse spectral
method has the property that t → q(·, t) ∈ C 1 ((−∞, 0) ∪ (0, ∞), C0 (R2 )). From
Theorem 5.1, Proposition 6.3 and its analog for t < 0 we have the following
theorem.
THEOREM 6.4. Let q0 be a Schwartz function which satisfies the smallness as-
sumption (2.16). The solution q(x, y, t) for (1.1) obtained by the inverse spectral
method is the unique generalized solution of the forward problem for (1.1) in
the class C([0, ∞), H 3 (R2 )) ∩ C 1 ((0, ∞), L∞ (R2 )). It is also the unique gen-
eralized solution of the backward problem in the class C((−∞, 0], H 3 (R2 )) ∩
C 1 ((−∞, 0), L∞ (R2 )).
Finally we prove a conservation law.
LEMMA 6.5. Let q(x, y, t) be the solution obtained by the inverse spectral method.
Then we have
Z Z
dx dy q (x, y, t) =
2
dx dy q02 (x, y). (6.18)
R2 R2
Proof. We use the notation in the proof of Proposition 6.3. From Equations (6.5)
and (6.6) (for qε and rε ) we have
Z Z
(d/dt) dx dy (qε ) = −3
2
dx dy (q 2 )ε (qε )x . (6.19)
R2 R2

Z
dx dy (qε )2 (x, y, t)
R2
Z
= dx dy (qε )2 (x, y, 0)−
R2
Z t Z
0
−3 dt dx dy (q 2 )ε (x, y, t 0 )(qε )x (x, y, t 0 ). (6.20)
0 R2
The conservation law (6.18) follows by letting ε → 0 in (6.20). 2
REMARK 6.6. When q0 is real, the uniqueness of solution for (1.1) implies that
q(x, y, t) is real for all t ∈ R and hence kq(x, y, t)kL2 (R2x,y ) is conserved.
24 LI-YENG SUNG
Acknowledgment
This work was supported in part by the National Science Foundation under Grant
DMS-94-96154.
References
1. Ablowitz, M. J. and Villarroel, J.: On the Kadomtsev–Petviashvili equation and associated
constraints, Stud. Appl. Math. 85 (1991), 195–213.
2. Boiti, M., Pempinelli, F. and Pogrebkov, A.: Solutions of the KPI equation with smooth initial
data, Inverse Problems 10 (1994), 505–519.
3. Boiti, M., Pempinelli, F. and Pogrebkov, A.: Properties of solutions of the Kadomtsev–
Petviashvili I equation, J. Math. Phys. 35 (1994), 4683–4718.
4. Faminskii, A. V.: The Cauchy problem for the Kadomtsev–Petviashvili equation, Russian Math.
Surveys 5 (1990), 203–204.
5. Fokas, A. S. and Ablowitz, M. J.: On the inverse scattering of the time-dependent Schrödinger
equation and the associated Kadomtsev–Petviashvili (I) equation, Stud. Appl. Math. 69 (1983),
211–228.
6. Fokas, A. S. and Sung, L.-Y.: The Cauchy problem for the Kadomtsev–Petviashvili-I equation
without the zero mass constraint, Math. Proc. Camb. Phil. Soc. 125 (1999), 113–138.
7. Johnson, R. S.: Water waves and Korteweg–de Vries equations, Fluid Mech. 97 (1980), 701–
719.
8. Kadomtsev, B. B. and Petviashvili, V. I.: On the stability of solitary waves in weakly dispersive
media, Soviet Phys. Dokl. 15 (1970), 539–541.
9. Lipovskij, V. D., Matveev, V. B. and Smirnov, A. O.: On the isomorphism between KP-equation
and Johnson equation, Zap. Semin. LOMI 150 (1986), 70–75.
10. Manakov, S. V.: The inverse scattering transform for the time-dependent Schrödinger equation
and Kadomtsev–Petviashvili equation, Physica D 3 (1981), 420–427.
11. Muskhelishvili, N. I.: Singular Integral Equations, Dover, New York, 1992.
12. Saut, J.-C.: Remarks on the generalized Kadomtsev–Petviashvili equations, Indiana Math. J.
42 (1993), 1011–1027.
13. Saut, J.-C.: Recent results on the generalized Kadomtsev–Petviashvili equations, Acta Appl.
Math. 39 (1995), 477–487.
14. Segur, H.: Comments on IS for the Kadomtsev–Petviashvili equation, in: M. Tabor and Y. M.
Treve (eds.), Mathematical Methods in Hydrodynamics and Integrability in Dynamical Systems,
AIP Conference Proceedings, No. 88, 1982, pp. 211–228.
15. Stein, E.: Singular Integrals and Differentiability Properties of Functions, Princeton University
Press, Princeton, 1970.
16. Tom, M. M.: On a generalized Kadomtsev–Petviashvili equation, Contemp. Math. 200 (1996),
193–210.
17. Ukai, S.: Local solutions of the Kadomtsev–Petviashvili equation, J. Fac. Sci. Univ. Tokyo,
Sect. IA Math. 36 (1989), 193–209.
18. Wloka, J. T.: Partial Differential Equations, Cambridge University Press, Cambridge, 1987.
19. Yosida, K.: Functional Analysis, Springer-Verlag, Berlin, 1995.
20. Zhou, X.: Inverse scattering transform for the time dependent Schrödinger equation with
applications to the KP-I equation, Comm. Math. Phys. 128 (1990), 551–564.
25
Soliton Asymptotics of Solutions of the

Sine-Gordon Equation
WERNER KIRSCH
Institute of Mathematics of the Ruhr-University Bochum, D-44780 Bochum, Germany
VLADIMIR KOTLYAROV
Mathematical Division, B. Verkin Institute for Low Temperature Physics, 310164 Kharkov, Ukraine
(Received: 26 November 1997; accepted in final form: 18 February 1999)

Abstract. An asymptotic analysis of the Marchenko integral equation for the sine-Gordon equation
is presented. The results are used for a construction of soliton asymptotics of decreasing and some
non-decreasing solutions of the sine-Gordon equation. The soliton phases are shown to have an
additional shift with respect to the reflectionless case caused by the non-zero reflection coefficient
of the corresponding Dirac operator. Explicit formulas for the phases are also obtained. The results
demonstrate an interesting phenomenon of splitting of non-decreasing solutions into an infinite series
of asymptotic solitons.
Mathematics Subject Classifications (1991): 35Q51, 35Q53, 81U40, 35B40.
Key words: solitons, sine-Gordon equation, Marchenko equation.
Introduction
It has been known for a long time [1 – 3] that the asymptotics of the solution of
the initial value problem with decreasing initial data is a superposition of solitons.
This superposition is the main term of the asymptotics. The next term tends to zero
when t → ∞. Sometimes, the phases of these solitons are the same as in the reflec-
tionless case. They depend on the eigenvalues and the normalization coefficients
only. This situation takes place for those nonlinear evolution equations integrated
by the inverse scattering transform, for which the kernel of the Marchenko integral
equation has no stationary point in the soliton domain. These are, for example,
the Korteweg–de Vries equation and the modified one. For many other integrable
equations, the kernel of the Marchenko integral equation has a stationary point. In
this case, the phases of solitons depend, in addition, on the reflection coefficient
[4 – 6]. It is important to remark that the additional phase shift, contributed by
non-zero reflection coefficient, has a finite order for t → ∞.
The first regorous results on this subject were obtained in the well-known papers
by A. B. Shabat [7] for the Korteweg–de Vries equation with decreasing initial
data, and by E. Ya. Khruslov for a step-like initial function [8]. Recently, Deift, Its
26 WERNER KIRSCH AND VLADIMIR KOTLYAROV
and Zhou [9, pp. 181–204] formulated the steepest descent method for the study
of long-time asymptotics for integrable nonlinear wave equations, based on the
oscillatory Riemann–Hilbert problem. This is a modern and very powerful method
for the asymptotic investigation of decreasing solutions of nonlinear wave equa-
tions as well as for the Painlevé equations, and for some models of quantum and
statistical physics. The references to this approach can be found in the book [9].
The Riemann–Hilbert method was also applied by Fokas and Its to the study of
initial boundary value problems on the semi-axis [10, 11]. Nevertheless, in the case
of non-decreasing solutions it is more convenient to use the associated Marchenko
integral equations.
The aim of this paper is twofold. The first is to present rigorous results on
the soliton asymptotics of a solution of the sine-Gordon equation with decreas-
ing and some non-decreasing initial data based on direct asymptotic analysis of
the Marchenko integral equation. The second is to obtain explicitly the additional
phase shift caused by the non-zero reflection coefficient.
Taking into account the equivalence [2] of the sine-Gordon equation given in
laboratory coordinates and in light-cone coordinates, we restrict our consideration
to the equation
uxt + 4 sin u = 0. (1)
A large class of solutions of Equation (1) can be constructed as follows. Let the
system of integral equations
Z ∞
K1 (x, y, t) + K2 (x, z, t)H (y + z, t) dz = 0 (x < y < ∞),
x
Z (2)
∞
K2 (x, y, t) + K1 (x, z, t)H (y + z, t) dz = −H (x + y, t),
x
where
Z
H (y, t) = eiλy+2it /λ dρ(λ), ⊂ C,

have a unique solution, which is sufficiently smooth. Then the function u(x, t)
defined by the equation
ux (x, t) = 4iK2 (x, x, t)
is a solution of Equation (1).

In what follows, we consider two cases. The first is the problem for Equation
(1) which we rewrite in the form
Z ∞

vt + 4 sin ∂x−1 v = 0, ∂x−1 v = v(s, t) ds, (3)
x
SOLITON ASYMPTOTICS OF SOLUTIONS OF THE SINE-GORDON EQUATION 27
with decreasing initial data

u(x, 0) = w(x), x ∈ R, w(x) → 0, |x| → ∞. (4)
In this case, the kernel H (y, t) coincides with (1.2) (see below). The second one is
the case for which
w(x) → 0, as x → ∞, w(x) → wn (x), as x → −∞, (5)
where wn (x) is an almost periodic function of a finite-gap type. The problem (3),
(5) was studied in [12] for the reflectionless case.
Let us formulate the main results of the present paper.
THEOREM 1. Let w(x) be a real Schwarz type function (w(x) ∈ S(R)). Then,
uniformly with respect to x ∈ R for t → ∞, the solution of the problem (3), (4)
has the asymptotic form
∂ Im det[I + D(x, t)]
v(x, t) = 4 arctan + o(1), t → ∞.
∂x Re det[I + D(x, t)]
The entries of the matrix D have the form
imj exp[2it/λj ]
Dj l (x, t) = exp i(λj x + θj (X) + λl x + θl (X)) ,
λj + λl
where
Z
1 α(X)
ln[1 + |r(s)|2 ] ds 1 x
θj (X) = , α(X) = √ , X= .
2π −α(X) s − λj 2X 2t
COROLLARY 1. The main term of the asymptotics of any decreasing solution

u(x, t) can be found by the same scheme as in the reflectionless case. To this end
we have to replace eiλj x with eiλj x+iθj (X) .
COROLLARY 2. The above mentioned solution has the form

X
n X
p
u(x, t) = vj (x, t) + wj (x, t) + o(1), t → ∞,
j =1 j =1
where

vj (x, t) = 4σj arctan exp −2νj x − νj−2 t − xj + 2π σj (j − 1),
σj = sign Im(mj )
is the soliton-kink with the phase
Z νj
1 ln[1 + |r(s)|2 ] ds
xj = xj −
0
, j = 1, 2, . . . , n,
2π −νj s 2 + νj2
and

Im λn+j sin[2 Re λn+j (x + |λn+j |−2 t + βj )]
wj (x, t) = 4 arctan
Re λn+j cosh[2 Im λn+j (x − |λn+j |−2 t − λj )]
is the soliton-breather with the phases
Re θn+j Im θn+j
βj = βj0 + , γj = γj0 − ,
Re λn+j Im λn+j
where
Z |λn+j |
1 ln[1 + |r(s)|2 ] ds
θn+j = , j = 1, 2, . . . , p.
2π −|λn +j | s − λj
The soliton phases xj0 , βj0 and γj0 have to be taken in the same form as in the
reflectionless case (r(λ) ≡ 0).
In the non-decreasing case, the kernel of the Marchenko integral equation is as
follows
Z ∞ 2n Z
1 X X N
1
H (y) = r(λ) e dλ +
iλy
h(λ) e dλ +
iλy
mj (t) eiλj y . (6)
2π −∞ 2π k=1 γk j =1
This form of the kernel corresponds to the following structure of the spectrum of
the Dirac operator with a “step-like” potential w(x): this operator has a continuous
spectrum of multiplicity two on the real line R of the λ-plane, and a continuous
spectrum of multiplicity one on a set of analytic arcs γj (j = 1, 2, . . . , 2n) with
end-points that are zeros of the polynomial
Y
k Y
m+k+1
P (λ ) =
2
λ −
2
Ei2 λ2 − Ej2 λ2 − E 2j ,
i=1 j =k+1
Re Ei = 0, i 6 k, Im Ej > 0, k + 2m = n,
which corresponds to the Riemann surface of an almost periodic potential wn (x)
[12] and, possibly, a finite number of eigenvalues λj which lie on the imaginary
axis of the λ-plane and symmetrically with respect to this axis.
Let σ be the complete spectrum of the Dirac operator
d w(x)
L = iσ3 +i σ2 ,
dx 2
which acts in the space L2 (R). Suppose that the condition
C = max |λ|−2 > 0 (7)
λ∈σ \R
bj ∈ σ \
holds and the maximal value is attained in a finite number of points E
R, j = 1, 2, . . . , 2p.
Let

1
DN = (x, t) : x > Ct − ln t N+1
, t > t (N) , (8)
2 Im E
bj .
where Im E = max16j 6n Im E
THEOREM 2. Let condition (7) be fulfilled and for all eigenvalues λj , |λj | > √1C .
Then for any natural number N, the solution of problem (3), (5) in the domain (8)
has the asymptotic form
∂ Im det[I + (ξ, t)]
v(x, t) = 4 arctan + o(1), ξ = x − Ct, t → ∞, (9)
∂x Re det[I + (ξ, t)]
p
where (ξ, t) = kB lr (ξ, t)kl,r=1 is a block matrix with the entries
X
N−k−1
wl(k+m) (t) X X lrpq
m j
lr
Bkj (x, t) = lr
G (t)Jp+q (ξ ),
m=0
k!m!t k+m+3/2 p=0 q=0 mj
Z ∞
b b
Jplr = τ p ei(El +Er )τ dτ.
ξ
lrpq
The functions ωl(n) (t) and Gmj (t) are determined by the scattering data of the
initial function w(x) and they are uniformly bounded with respect to t.
THEOREM 3. Under the conditions of Theorem 2, let us suppose that the max-
b1 = ia and E
imal value (7) is attained in two points E b2 = −ia, a > 0. Then
formula (9) takes the form
[(N+1)/2]
∂ X
v(x, t) = 2π(n − 1) +
∂x n=1

+ 4 arctan exp −2a x − a −2 t − αn (x, t) + o(1), t → ∞, (10)
where
Z √
1 1 t /x
ln[1 + |r(µ)|2 ] dµ
αn (x, t) = αn0 − ln t 2n−1/2 − √
. (11)
2a 2π − t /x µ2 + a 2
The numbers αn0 are determined (Equation (5.11)) by the scattering data of the
initial function w(x), and r(µ) is the reflection coefficient of the Dirac operator L.
THEOREM 4. Under the conditions of Theorem 2, let us suppose that the maxi-
mal value in (7) is attained in four points E b1,2 = ±E, Eb3,4 = ±E. Then formula
(9) takes the form
[(N+1)/2]
∂ X Im E sin[2 Re E(x + |E|−2 t + βn (x, t))]
v(x, t) = 4 arctan +
∂x n=1 Re E cosh[2 Im E(x − |E|−2 t + γn (x, t))]
+ o(1), (12)
Z √
1 t /x
ln[1 + |r(µ)|2 ] dµ
βn (x, t) = βn0 + Re √
, (13)
2π Re E − t /x µ−E
1
γn (x, t) = γn0 + ln t 2n−1/2 −
2 Im E
Z √t /x
1 ln[1 + |r(µ)|2 ] dµ
− Im √
. (14)
2π Im E − t /x µ−E
The numbers βn0 , γn0 are determined (Equations (5.12), (5.13)) by the scattering
data of the initial function w(x).
REMARK. √ If the Dirac operator has a discrete spectrum which lies inside the circle
of radius 1/ C, then the superposition of usual solitons, i.e. kinks (Re λk = 0) and
breathers (Re λk 6= 0), should be added to Equation (10) and (12) of asymptotic
solitons. Each asymptotic soliton gets an additional phase shift caused by solitons
generated by the discrete spectrum.
1. Representation of the Solution

Due to the inverse scattering transform, the solution of the problem (3), (4) is
represented as follows [1, 2]
Z ∞
2
v(x, t) = 4iK2 (x, x, t), v(s, t) ds = 8K1 (x, x, t), (1.1)
x
where K1 (x, y, t) = K 1 (x, y, t), K2 (x, y, t) = −K 2 (x, y, t), and these functions
solve the integral Marchenko equation (2). For this case, the kernel of Equation (2)
has the form
X N Z ∞
1
H (y, t) = mj (t) eiλj y
+ r(λ, t) eiλy dλ (1.2)
j =1
2π −∞
with
mj (t) = mj e2it /λj , r(λ, t) = r(λ) e2it /λ .
Here the numbers λj (j = 1, 2, . . . , N) are the eigenvalues of the Dirac operator
L. Since w(x) is a real function, the eigenvalues are distributed in two subsets [2]:
λj = iνj , νj > 0, j = 1, 2, . . . , n;
λn+p+j = −λ̄n+j , Re λn+j > 0, Im λn+j > 0, j = 1, 2, . . . , p.
Therefore N = n + 2p. The coefficients mj (j = 1, 2, . . . , N) corresponding to

the eigenvalues λj have the following properties:
mj = −mj , j = 1, 2, . . . , n;
mn+p+j = −mn+j , mn+j ∈ C, j = 1, 2, . . . , p.
The reflection coefficient r(λ) satisfies the equation
r(−λ) = −r̄(λ), λ ∈ R.
So the kernel (1.2) possesses the property
H (y, t) = −H (y, t). (1.3)
Let Hb be an integral operator which acts in the space L(i) [x, ∞) (i = 1, 2, ∞)

of vector-functions f (y) = (f1 (y), f2 (y)) according to the formula
Z ∞
b 0 H (y + z, t) f1 (y)
H f (y) = dz.
x H (y + z, t) 0 f2 (y)
Then the Marchenko integral equation (2) can be written in the form

I +Hb K = G, K = (K1 , K2 ), G = (0, −H ). (1.4)
It is known [2] that for any t ∈ R Equation (1.4) has a unique solution in each
space L(i) [x, ∞) (i = 1, 2, ∞) and if r(λ) ∈ S(R), then K(x, y, t) is a Schwarz
type vector-function for each t, too.
2. Degenerate Integral Equation

Let us write the kernel (1.2) in the form
H (y, t) = HN (y, t) + R(y, t)
and let L = (L1 (x, y, t), L2 (x, y, t)) be the solution of the integral equation

I +Rb L = GR , GR = (0, −R). (2.1)
Since the operator R b is

b is compact in any space L(i) [x, ∞) (i = 1, 2, ∞) and R
skew-symmetric (1.3) in the space L [x, ∞), Equation (2.1) has a unique solution
2
in each space L(i) [x, ∞) (i = 1, 2, ∞) for every t ∈ R. Therefore

I +Rb −1 6 1
2
and it is easy to obtain an estimate, uniform with respect to x, t, for the vector-
function L
Z ∞

L(x, ., t) 2 6 kGR k2 = kRk2 = σ (2x, t) = R(s, t)2 ds
2 2 2
Z ∞ 2x
1 2
6 r(λ) dλ = C 2 , (2.2)
2
2π −∞
where
Z
∞ 2
L(x, ., t) 2 = |L1 (x, y, t)|2 + L2 (x, y, t) dy.
2
x
Moreover, there is a uniform estimate in the space L∞ [x, ∞) for the components
Lj (x, y, t)

L1 (x, ., t) 6 kRk2 kL2 k2 6 σ (2x, t) 6 C 2 , (2.3)
∞ 2

L2 (x, ., t) 6 kRk∞ + kRk2 kL1 k2
∞
6 sup |R(y, t)| + σ (2x, t) 6 C1 + C22 , (2.4)
y>2x
where
Z ∞
1
C1 = |r(λ)| dλ.
2π −∞
Let us find the solution of Equation (1.4) in the form

K =L+ I +Q b M, (2.5)
b satisfies the equation
where an integral operator Q

I +Q b I +R b = I +R b I +Q b = I. (2.6)
LEMMA 1. The operator Q b is a bounded integral operator, and the correspond-

ing kernel has the form

Q1 (y, z) Q2 (y, z)
Q(y, z) = ,
Q2 (y, z) Q1 (y, z)
where
( Rz
L1 (z, y) + x [L1 (s, y)L1 (s, z) + L2 (s, y)L2 (s, z)] ds, z < y,
Q1 (y, z) = Ry
L1 (y, z) + [L1 (s, y)L1 (s, z) + L2 (s, y)L2 (s, z)] ds, z > y,
( Rxz
L2 (z, y) + x [L2 (s, y)L1 (s, z) + L1 (s, y)L2 (s, z)] ds, z < y,
Q2 (y, z) = Ry
L2 (y, z) + x [L2 (s, y)L1 (s, z) + L1 (s, y)L2 (s, z)] ds, z>y
and
Q1 (y, z) = Q1 (z, y) = Q1 (z, y), Q2 (y, z) = Q2 (z, y) = −Q2 (z, y).
Proof. The operator identities (2.6) are equivalent to pair equations of Marchenko
type on functions Q1 (y, z) and Q2 (y, z). Since the functions L1 (y, z) and L2 (y, z)
are the solution of Equation (2.1), it is easy to check that the representations for
Q1 (y, z) and Q2 (y, z) are valid. 2
Let us introduce some notations:

Z ∞
(1)
hj (x, t) = e iλj x
+ L1 (x, s, t) eiλj s ds,
Z ∞ x
h2j (x, t) = L2 (x, s, t) eiλj s ds.

x
Z y
gj(1) (x, y, t) = hj(1)(y, t) + hj(1)(s, t)L1 (s, y, t) +
x

+ hj(2)(s, t)L2 (s, y, t) ds,
Z y
(1)
gj(2) (x, y, t) = hj(2)(y, t) + hj (s, t)L2 (s, y, t) +
x

+ hj(2)(s, t)L1 (s, y, t) ds.
The next lemma presents a degenerate integral equation for the problem (3), (4).
LEMMA 2. The vector-function M(x, y, t) is the solution of the equation

bN M = −GN .
M +F (2.7)
The kernel of this equation is degenerate and has the form
(1)
FN (x, y, z, t) FN(2) (x, y, z, t)
FN (x, y, z, t) = ,
FN(2) (x, y, z, t) FN(1) (x, y, z, t)
where
X
N
FN(1) (x, y, z, t) = mj (t) eiλj y gj(2)(x, z, t),
j =1
(2.8)
X
N
FN(2) (x, y, z, t) = mj (t) eiλj y gj(1)(x, z, t).
j =1
The vector-function GN (x, y, t) is given by the relations
(1)
X
N
GN (x, y, t) = FN(1) (x, y, x, t) = mj (t) eiλj y h(2)
j (x, t),
j =1
(2.9)
(2)
X
N
GN (x, y, t) = FN(2) (x, y, x, t) = mj (t) eiλj y h(1)
j (x, t).
j =1
Proof. It follows from Equations (1.4), (2.1), (2.5) and (2.6) that

M +H bN I + Q b M = −H bN L + GHN .
bN = H
Then the operator F bN (I + Q)
b is degenerate and the corresponding kernel has
the form (2.8). The vector-function GN = −H bN L + GHN and the corresponding
components have the form (2.9). 2
LEMMA 3. The solution of Equation (2) with kernel (1.2) has the representation
X
N

K1 (x, y, t) = L1 (x, y, t) + Xj (x, t)gj(1) (x, y, t) +
j =1

+ Yj (x, t)gj(2) (x, y, t) ,
(2.10)
X
N

K2 (x, y, t) = L2 (x, y, t) + Xj (x, t)gj(2) (x, y, t) +
j =1

+ Yj (x, t)gj(1) (x, y, t) ,
where L = (L1 , L2 ) is the solution of Equation (2.1) and the N-dimensional vec-
tors X = (X1 , X2 , . . . , XN ) and Y = (Y1 , Y2 , . . . , YN ) are the solution of the
system of linear algebraic equations
X
N
Xj + (Aj l Xl + Bj l Yl ) = −mj (t)hj(2) ,
l=1
(2.11)
X
N
Yj + (Aj l Yl + Bj l Xl ) = −mj (t)hj(1) .
l=1
The entries of the matrices A and B have the form

Z ∞
(1)
Aj l (x, t) = mj (t) hj (s, t)hl(2) (s, t) + hj(2) (s, t)hl(1)(s, t) ds,
x
Z ∞ (2.12)
(1) (1) (2) (2)
Bj l (x, t) = mj (t) hj (s, t)hl (s, t) + hj (s, t)hl (s, t) ds.
x
Proof. Let us put
X
N X
N
M1 (x, y, t) = iλj y
Xj (x, t) e , M2 (x, y, t) = Yj (x, t) eiλj y .
j =1 j =1
Then Equation (2.7) is equivalent to the system (2.11) of linear algebraic equations,
and Equation (2.5) leads to the representation (2.10). 2
COROLLARY 3. The solution of the problem (3), (4) has the form
X
N

v(x, t) = 4iL2 (x, x, t) + 4i Xj (x, t)h(2) (1)
j (x, t) + Yj (x, t)hj (x, t) .
j =1
So, the solution of the initial value problem (3), (4) is completely defined via
the solution L of Equation (2.1) and Jost type functions hj(1)(x, t) and hj(2)(x, t).
Therefore we have to study the asymptotic behaviour of these functions for a large
time.
3. Asymptotic Analysis of the Marchenko Equation

In what follows it is convenient to introduce new (“slow”) variables X = x/2t,
Y = y/2t, Z = z/2t, etc. The kernel R(y, t) has a stationary point. Therefore one
can obtain both the representation
R(y, t) = R0 (y, t) + R1 (y, t),
where

1 h √ √ i e iπ4 −3/4 1
R0 (y, t) = √ r̂(Y ) e4it Y − r̂ Y e−4it Y , r̂(Y ) = √ Y r √ ,
t 8π Y
and the estimate

R1 (y, t) 6 Ct −3/2 ρ(Y ), (3.1)
where ρ(Y ) is the absolute value of a function of the Schwarz type.
Let us look for the solution of Equation (2.1) in the form
L = P + S. (3.2)
If a vector-function P is the solution of the equation

I +R b0 P = GR0 (3.3)
then the vector-function S has to satisfy the equation

I +Rb S = GR1 − R b1 P . (3.4)
LEMMA 4. The estimates

C C
kSk2 6 √ , kSk∞ 6 √ (3.5)
t t
are valid for the vector-function S.
Proof. It follows from Equation (3.4) that

kSk2 6 I + R b −1 GR1 − R b1 kP k2 .
b1 P 6 kGR1 k2 + R
2 2 2
Estimate (3.1) implies that

Z ∞ 1/2
C C(ρ)
kGR1 k2 = kR1 k2 6 2 2
ρ (Y ) dY = , (3.6)
t X t
Z ∞ 1/2

b1 6 2C
R √ 2 Yρ (Y + 2X) dY
2
2
t 0
C(ρ, X0 )
6 √ , X0 < X < ∞ (X0 > −∞). (3.7)
t
We find from Equation (3.3) and estimates (3.6), (3.7) that
Z ∞ 1/2
2

kP k2 6 kGR0 k2 = kR0 k2 6 r̂(Y ) 2 =
r̂(Y ) dY b2 . (3.8)
6C
X
So, the first estimate (3.5) is valid. For the second one we can write

kSk∞ 6 kGR1 k∞ + kR1 k2 kP k2 + R b kSk2 .
2
By using Equations (3.1), (3.6), (3.8), (2.2) and taking into account the first esti-
mate (3.5) we finally obtain
C b2
C(ρ)C C
kSk∞ 6 sup |ρ(Y )| + + C(ρ, X0 ) √ .
t 3/2 Y >X t t
COROLLARY 4. Estimates (2.3) and (2.4) imply that kP k∞ is uniformly bounded

with respect to x and t. Moreover, kP1 k∞ = O(1) and kP2 k∞ = O( √1t ) for t → ∞.
Let us find the solution P = (P1 , P2 ) in the form

P1 (x, y, t) = N1 (X, Y, t),
1 h i
√ √
(3.9)
P2 (x, y, t) = √ N2 (X, Y, t) e4it X+Y − N 2 (X, Y, t) e−4it X+Y .
t
Then we obtain the representation
Z ∞h √ √ i
N1 (X, Y, t) = −2 N2 (X, Z, t)r̂(Z + Y ) e4it ( X+Z+ Z+Y ) + c.c. dZ +
Z ∞h X
√ √ i
+2 ¯ + Y ) e−4it ( X+Z+ Z+Y ) + c.c. dZ
N2 (X, Z, t)r̂(Z (3.10)
X
for N1 (X, Y, t) and the integral equation for N2 (X, Y, t)
Z ∞ √ √
N2 (X, Y, t) = N2 (X, S, t) e4it ( X+S− X+Y ) 0(X, Y, S, t) dS −
X
Z ∞ √ √
− N 2 (X, S, t) e−4it ( X+S+ X+Y ) 0(X, Y, S, t) dS
X
= −r̂(X + Y ). (3.11)
The kernel of this equation is

Z ∞ √ √
0(X, Y, S, t) = 4t ¯ + S) e4it ( Y +Z− Z+S) dZ −
r̂(Y + Z)r̂(Z
X
Z ∞ √ √
− 4t r̂(Y + Z)r̂(Z + S) e4it ( Y +Z+ Z+S) dZ
X
= 01 + 02 .
Introduce the operators
Z ∞ √ √
b0t f (Y ) =
1
e4it ( X+S− X+Y ) 01 (X, Y, S, t)f (S) dS,
ZX∞ √ √
b0t2 f (Y ) = e4it ( X+S− X+Y ) 02 (X, Y, S, t)f (S) dS,
X
Z ∞ √ √
(3.12)
b0t3 f (Y ) = e−4it ( X+S+ X+Y ) 02 (X, Y, S, t)f (S) dS,
ZX∞ √ √
b0t4 f (Y ) = e−4it ( X+S+ X+Y ) 01 (X, Y, S, t)f (S) dS.
X
Then Equation (3.11) and the complex conjugated one may be written as follows

N2 + b0t1 + b
0t2 N2 − b 0t3 + b
0t4 N 2 = −r̂,
4 2
N2 + b 0t + b0 t N2 − b 0t + b
3 1
¯
0 t N 2 = −r̂,
i.e.,
f +b0t f = g, (3.13)
b1 b2
0t −b0t3 0t −b0t4
b
0t = + . (3.14)
b
−0t
3
b 1
0t −b0t
4
b0
2
t
It is easy to see that
∗ b1 ∗ ∗
0t = b =b
1 3
0t1 = b
b 0t1 , 0t , b
0t3 0t .
Therefore the first summand in (3.14) is a self-adjoint operator in L2 [X, ∞).
The second summand is not self-adjoint in the same space.
LEMMA 5. For any X > X0 > −∞ and for any fixed t, the operator b 0t is a
compact operator in the space L2 [X, ∞), and Re(b 0t f, f ) > 0 for any vector-
function f (Z) ∈ L2 [X, ∞) and for sufficiently large t.
Proof. The compactness of the operator b0t follows from the fact that it is an op-
erator of the Hilbert–Schmidt type in the space L2 [X, ∞). This statement follows
from the inequality
Z ∞
2
Y r̂(Y + 2X) dY < ∞. (3.15)
0
Since r(λ) ∈ S(R), (3.15) holds. Moreover, we find that

Z ∞

b
0t f, f = 4t F1 (X, Z, t) + F2 (X, Z, t)2 dZ −
X
Z ∞

− 4t F1 (X, Z, t) + F2 (X, Z, t) G1 (X, Z, t) +
X

+ G2 (X, Z, t) dZ,
where
Z ∞ √ √
F1 (X, Z, t) = r̂(Z + Y )f¯1 (Y ) e4it ( Y +Z− X+Y )
dY,
ZX∞ √ √
F2 (X, Z, t) = ¯ + Y )f¯2 (Y ) e4it (
r̂(Z Y +Z− X+Y )
dY,
ZX∞ √ √
Y +Z+ X+Y )
G1 (X, Z, t) = r̂(Z + Y )f1 (Y ) e4it ( dY,
ZX∞ √ √
G2 (X, Z, t) = ¯ + Y )f2 (Y ) e−4it (
r̂(Z Y +Z+ X+Y )
dY.
X
By virtue of condition (3.15) these functions belong to the space L2 [X, ∞) in
the variable Z, and Fj (X, Z, t) = O(1) and Gj (X, Z, t) = o(1), for t → ∞.
Therefore

Re b 0t f, f > 4tkF1 + F2 k kF1 + F2 k − kG1 + G2 k > 0.
The lemma is proved. 2
COROLLARY 5. For sufficiently large t, Equation (4.14) has a unique solution

in the space L2 [X, ∞), and
−1
I +b0t 2 6 1.
LEMMA 6. For any function ϕ(Y ) ∈ D[X, ∞) = L2 [X, ∞) ∩ C 1 [X, ∞) and

operators (3.12), the relations
lim b
0ti ϕ = 0, i = 2, 4,
t →∞
 Z ∞

 X + Y 1/4 X + S 1/4

 − + ×

 X+S X+Y


X

 r0 (X + Y )r̄0 (X + S)


 × 4π i(S − Y + i0) ϕ(S) dS, i = 1,
lim b
0i ϕ = Z ∞ (3.16)
t →∞ t 
 X + Y 1/4 X + S 1/4

 − − ×



 X+S X+Y


X

 r (X + Y )r0 (X + S)
× 0 ϕ(S) dS, i = 3,
4π(S − Y )
√
holds with r0 (Y ) = r(1/ Y ) and r(λ) the reflection coefficient; the equalities are
valid in the metric of the space L2 [X, ∞).
This lemma can be proved as in [13].

Let
b1 3
0∞ −b 0∞
b
0∞ =
−b 0∞ b
3 1
0∞
be an operator in the space of vector-functions L2 [X, ∞), which is determined by
the r.h.s. of Equation (3.16). It is easy to check that this operator is self-adjoint and
nonnegative. Therefore the operator (I + b 0∞ )−1 exists, and k(I + b 0∞ )−1 k 6 1.
LEMMA 7. For any vector-function f (Y ) ∈ L2 [X, ∞),

−1 −1
I +b
0t f − I + b 0∞ f = o(1), as t → ∞.
Proof. Let ϕ(Y ) ∈ D[X, ∞), h = (I + b

0∞ )ϕ . Then the value
−1 −1 −1
I +b 0t h 6 I + b
0∞ h − I + b 0t b 0∞ ϕ
0t ϕ − b

6 b 0∞ ϕ
0t ϕ − b
tends to zero as t → ∞ by virtue of Lemma 6. The set of functions h = (I + b

0∞ )ϕ
is the dense set in the space L2 [X, ∞), the operator (I + b0∞ )−1 exists and is
bounded. Hence, due to the resolvent convergence [14],
−1 −1
I +b
0t f → I +b
0∞ f, ∀f ∈ L2 [X, ∞),
Lemma 7 is proved. 2
It follows from the given proof that for the solution N2 (X, Y, t) of Equation
(3.11) the following estimate is valid

N2 (X, Y, t) − N(X, Y ) = o(1), t → ∞, (3.17)
2
where the function N(X, Y ) satisfies the equation

Z ∞
X + Y 1/4 X + S 1/4
N(X, Y ) − N(X, S) + ×
X X+S X+Y
Z ∞
r0 (X + Y )r̄0 (X + S) X + Y 1/4 X + S 1/4
× ds + N (X, S) − ×
4π i(S − Y + i0) X X+S X+Y
r0 (X + Y )r0 (X + S)
× dS = −r̂(X + Y ). (3.18)
4π(S − Y )
LEMMA 8. Equation (3.18) has a unique solution in the space L2 [X, ∞) which
can be represented in the explicit form
N(X, Y ) = −r̂(X + Y ) ×
Z ∞
i X + Y 1/2 ln[1 + |r0 (X + S)|2 ]
× exp dS . (3.19)
2π X X+S S − Y + i0
The function N(X, Y ) is bounded for Y ∈ [X, ∞).
Proof. We look for the solution N(X, Y ) in the form
N(X, Y ) = −r̂(X + Y )A(X, Y ),
where
eiπ/4
r̂(X + Y ) = √ (X + Y )−3/4 r0 (X + Y ).
2π
Then
Z r r !
∞
X+Y X + Y |r0 (X + S)|2 A(X, S)
A(X, Y ) − 1+ dS −
X X+S X+S 4π i(S − Y + i0)
Z ∞r r !
X+Y X + Y |r0 (X + S)|2 Ā(X, S)
− 1− dS = 1.
X X+S X+S 4π i(S − Y )
Let us take λ = (X + Y )−1/2 , µ = (X + S)−1/2 , B(X, λ) = A(X, 1/(λ2 − X)),

then we obtain
Z α Z α
|r(µ)|2 B(X, µ) dµ |r(µ)|2 B(X, µ) dµ
B(X, λ) = 1 − + ,
0 2π i(µ − λ − i0) 0 2π i(µ + λ)
q
where α = √12X = xt . In the last summand we take µ = −τ and

B(X, λ), λ > 0,
B0 (X, λ) =
B(X, −λ), λ < 0.
Then
Z α
|r(µ)|2 B0 (X, µ) dµ
B0 (X, λ) = 1 − , −α < λ < α.
−α 2π i(µ − λ − i0)
Due to the Riemann–Hilbert problem we find that
Z α
i ln[1 + |r(µ)|2 ] dµ
B0 (X, λ) = exp .
2π −α µ − λ − i0
If we come back to the initial variable, we obtain
( Z ∞ )
i X + Y 1/2 ln[1 + |r0 (X + S)|2 ]
A(X, Y ) = exp dS .
2π X X+S S − Y + i0
Since A(X, Y ) ∈ L∞ [X, ∞) is a continuous function with respect to Y 6= X, N(X,

Y ) is bounded for Y ∈ [X, ∞), belongs to the space L2 [X, ∞), and is a continuous
function with respect to Y 6= X. This lemma is proved. 2
Let us introduce new functions

Z ∞h √ √ i
L1 (x, y, t) = −2
0
N(X, Z)r̂(Z + Y ) e4it ( X+Z+ Z+Y ) + c.c dZ +
ZX ∞ h √ √ i
+2 ¯ + Y ) e4it ( X+Z− Z+Y ) + c.c. dZ,
N(X, Z)r̂(Z
X
1 h √ √ i
L02 (x, y, t) = √ N(X, Y ) e4it X+Y − N (X, Y ) e−4it X+Y , (3.20)
t
where N(X, Y ) is determined by formula (3.19). These functions belong to the
space L2 [X, ∞) ∩ L∞ [X, ∞), and L01 (x, y, t) is continuous with respect to Y ∈
[X, ∞), and L02 (x, y, t) is continuous with respect to Y 6= X, Y ∈ [X, ∞).
LEMMA 9. Let L1 (x, y, t), L2 (x, y, t) be the solution of Equation (2.1), and let
L1(0)(x, y, t), L2(0)(x, y, t) be defined by Equation (3.20). Then the estimates

L1 (x, y, t) − L0 (x, y, t) = o(1), t → ∞, (3.21)
∞
1

L2 (x, y, t) − L (x, y, t) = o(1), t → ∞,
0
(3.22)
2 2
are valid.
Proof. It follows from Equations (3.2), (3.9), (3.10) and estimates (3.5), (3.17)
that

L2 (x, . , t) − L(0)(x, ., t)
2 L2 [x,∞)

6 2 N2 (x, ., t) − N(x, ., t)
+ S2 (x, ., t)
L2 [X,∞) L2 [x,∞)
= o(1), t → ∞.
Estimate (3.22) can be easily deduced from Equation (3.11). 2
4. Soliton Asymptotics of the Solution

The further analysis will be connected with the study of the asymptotic behaviour
of the functions h1j (x, t), h2j (x, t), Aj l (x, t) and Bj l (x, t).
LEMMA 10. Let X > X0 > −∞. Then for t → ∞

h1j (x, t) = exp iλj x + iθj (X) 1 + εj (X, t) , (4.1)
where
Z
1 α
ln[1 + |r(µ)|2 ] dµ
θj (X) = B0 (X, λj ) = ;
2π −α µ − λj
1
h2j (x, t) = √ ηj (X, t) exp(iλj x). (4.2)
t
The functions εj (X, t) = o(1) as t → ∞, and the functions ηj (X, t) are uniformly
bounded with respect to X and t.
Proof. For h1j (x, t), by virtue of (3.20) and (3.21), we have
Z ∞
−iλj x
1
hj (x, t) e −1 = L1(0)(x, y, t) eiλj (y−x) dy −
x
Z ∞
(0)
− L1 (x, y, t) − L1 (x, y, t) eiλj (y−x) dy
Z ∞x
= L1(0)(x, y, t) eiλj (y−x) dy + o(1), t → ∞.
x
For the main integral, we write the relation

Z ∞
L1(0)(x, y, t) eiλj (y−x) dy
x
Z √
∞
e−8it X+Y dZ
= 4t r (X + Z)A(X, Z)
2
√ + c.c. −λ̄j +
x 2it (1/ X + Y − λj )
Z ∞
2 dZ
+ 4t r̂(X + Z) A(X, Z) √ + c.c. −λ̄j + O t −1
x 2it (λj − 1/ X + Y )
1 2
= J 1 (λj ) + J − λ̄j + J 2 (λj ) + J −λ̄j + O t −1 .
The function A(X, Z) does not have any limit as Z → X, namely
"r #
X+Y
A(X, Z) = exp i ν(X) ln(Z − X) A0 (X, Z),
X
where ν(X) = 2π 1
ln[1 + |r0 (2X)|2 ], and the function A0 (X, Z) is continuous and
bounded when Z → X. By using the last equality, it is easy to obtain the estimate
1
J 1 (λj ) + J −λ̄j = O t −1 , t → ∞.
2
The integrals J 2 (λj ) and J (−λ̄j ) give us
Z α Z α
2 |r(µ)|2 B(X, µ) dµ |r(µ)|2 B(X, µ) dµ
J (λj ) + J −λ̄j = −
2
+ ,
0 2π i(µ − λj ) 0 2π i(µ + λj )
if we use the replacement µ = (Y + Z)−1/2. Hence

h1j (x, t) = exp(iλj x)B0 (X, λj ) 1 + εj (X, t) , t → ∞.
Thus formula (4.1) is proved. 2
For h2j (x, t) we write

√ −iλ x 2
t e j hj (x, t)
Z ∞h √ √ i
= 2t N2 (X, Y, t) e4it X+Y − N 2 (X, Y, t) e−4it X+Y e2it λj (Y −X) dY +
X
Z ∞
√
+ t S2 (x, y, t) eiλj (y−x) dy.
x
Since kN2 (X, Y, t)k∞ is uniformly bounded (Corollary 4) with respect to X, Y, t,

and since estimate (3.5) holds, we come to Equation (4.2) with some functions
ηj (X, t) for which the estimate
C
ηj (X, t) <
Im λj
is valid. The lemma is proved.
Bearing in mind the previous lemmas, it is easy to prove the following
LEMMA 11. For the entries (2.13) of the matrices A and B, the relations
imj (t) ϕj l (X, t)
Aj l (x, t) = √ exp(i(λj + λl )x),
t λj + λl
and
imj (t)[1 + εj l (X, t)]
Bj l (x, t) = exp i(λj x + θj (X) + λl x + θl (X))
λj + λl
are true, and the functions ϕj l (X, t) are uniformly bounded with respect to X and
t.
From the previous results, we deduce the following statement.
THEOREM 10 . For t → ∞ the solution of the problems (3), (4) has the asymptotic
form
X
N
v(x, t) = 4i Yj(0)(x, t) eiλj x+iθj (X) + o(1), x > X0 t (X0 > −∞), (4.3)
j =1
where the functions Yj(0)(x, t) (together with Xj(0) ) solve the system of linear alge-
braic equations
X
N
Xj(0) (x, t) + Dj l (x, t)Yl(0) (x, t) = 0,
l=1
(4.4)
X
N
Dj l (x, t)Xl(0) (x, t) + Yj(0)(x, t) = −mj (t) eiλj x+iθj (X)
l=1
with the entries

imj (t)
Dj l (x, t) = exp i(λj x + θj (X) + λl x + θl (X)) ,
λj + λl
where
Z α(X)
1 ln[1 + |r(s)|2 ] ds
θj = .
2π −α(X) s − λj
Proof. Due to Lemma 11, it is natural to look for the solution of Equation (2.11)
in the form
1 1
X = X(0) + √ X(1) , Y = Y (0) + √ Y (1) .
t t
By using Equation (2.10) we find that Equation (4.3) together with (4.4) hold. As
in the reflectionless case one can obtain the determinant formula (4).
We have obtained the estimates uniformly with respect to X > X0 > −∞.
Therefore, it is necessary to consider the integral Marchenko equation in which the
variable y ∈ (−∞, x], and then we obtain the estimates uniformly with respect to
X 6 X0 < ∞.
Theorems 1 and 10 are proved. 2
5. Non-Decreasing Initial Problem

The reflectionless case of this problem was studied in [12]. Let x = Ct + ξ, y =
Ct + η, where the constant C is defined by Equation (7). In what follows, we
restrict our consideration to the domain DN (Equation (8)). It was shown [12] that
for t → ∞, the kernel (6) in the domain DN has the asymptotic form
H (x + y, t) = HN (x + y, t) + RN (x + y, t), (5.1)
where
X
p
X
N−1
(ξ + η)n ωk(n) (t)
HN (x + y, t) = exp[iEk (ξ + η)]
k=1 n=0
t n+3/2
X
p
X
N−1 X
N−k−1
= eiEl (ξ +η) ηk (l)
ωkm (t)ξ m , (5.2)
l=1 k=0 m=0
(l) ωl(k+m) (t)

ωkm (t) = ,
k!m!t k+m+3/2
and
RN (x + y, t) = R0 (x + y, t) + RN1 (x + y, t). (5.3)
The functions ωk(m) (t) were defined in [12]. It is important to remark here that they
are uniformly bounded with respect to t. The function R0 (y, t) is the same as that
in Equation (3.1), i.e.,

1 h √ √ i
−4it Y
e iπ4 −3/4 1
R0 (y, t) = √ r̂(Y ) e 4it Y
− r̂(Y ) e , r̂(Y ) = √ Y r √ .
t 8π Y
For the function RN1 (y, t), the estimate
1 N − Im Eη
R (y, t) 6 C1 ρ(Y ) + C2 + C3 |η| e (5.4)
N
t 3/2 t 2y2 t N+3/2
holds. Therefore, if L = (L1 , L2 ) is the solution of Equation (2.1) with kernel
(5.3), then for L = P + S we obtain now the estimates

1 ln t
kP1 k∞ = O(1), kP2 k∞ = O √ , kSk2,∞ = ON √ (5.5)
t t
as t → ∞.
Due to Equations (5.1)–(5.3) and estimates (5.4), (5.5), we are in the same
situation as in the previous case. But, since the degenerate kernel (5.2) has a form
different from (1.2), we need to introduce new functions of Jost type. Namely,
Z ∞
hn (x, t) = (x − Ct) e
l1 n iEl (x−Ct )
+ L1 (x, y, t)(y − Ct)n eiEl (y−Ct ) dy,
x
Z ∞ (5.6)
hn (x, t) =
l2
L2 (x, y, t)(y − Ct) e n iEl (y−Ct )
dy,
x
where L = (L1 , L2 ) is the solution of Equation (2.1) with the kernel (5.3).
Instead of Lemma 3 we obtain the following statement:
LEMMA 12. The solution of the initial value problem (3), (5) has the representa-
tion
X X
p N−1
v(x, t) = 4iL2 (x, x, t) + 4i n (x, t) +
Xnl (x, t)hl2
l=1 n=0

+ Ynl (x, t)hl1
n (x, t) , (5.7)
where the functions Xnl (x, t), Ynl (x, t) are the solution of the system of linear alge-
braic equations
X X
p N−1
X X
p N−1
Xkl + Alr r
ks Xs + Ys = akl ,
lr r
Bks l = 1, 2, . . . , p, (5.8)
r=1 s=0 r=1 s=0
X X
p N−1
X X
p N−1
Bks Xs + Ykl +
lr r
ks Ys = bk ,
Alr r l
k = 0, 1, . . . , N − 1.
r=1 s=0 r=1 s=0
The coefficients of this system are as follows.

X Z ∞
N−k−1
(l) l1
Alr
ks (x, t) = ωkm (t) s (y, t) + hm (y, t)hs (y, t) dy,
hm (y, t)hr2 l2 r1
m=0 x
X Z
N−k−1
(l)
∞
lr
Bks (x, t) = ωkm (t) m (y, t)hs (y, t) + hm (y, t)hs (y, t) dy,
hl1 r1 l2 r2
m=0 x
X
N−k−1
(l)
akl =− ωkm (t)hl2
m (x, t), (5.9)
m=0
X
N−k−1
(l)
bkl = − ωkm (t)hl1
m (x, t).
m=0
Since estimates (5.4) and (5.5) hold, Lemmas 5–9 are true under the restriction
that x, t ∈ DN . For the Jost type function (5.6), instead of Lemma 10 we obtain
the following one.
LEMMA 13. Let x, t ∈ DN . Then for t → ∞

X
k
hl1
k (x, t) = exp(iEl ξ ) hl1 m
km (X, t)ξ , ξ = x − Ct,
m=0
X
k
−1/2
k (x, t) = t
hl2 hl2 m
exp(iEl ξ ) km (X, t)ξ ,
m=0
where
k! ∂ k−m
km (X, t) = (−i)
hl1 +
k−m l
B0 (X, El ) 1 ε (X, t) ,
m!(k − m)! ∂Elk−m km
" Z √ #
i 1/ 2X
ln[1 + |r(µ)|2 ] dµ
B0 (X, El ) = exp √ , X = x/2t.
2π −1/ 2X µ − El
l
The function εkm (X, t) = o(1), as t → ∞, and the functions hl2
km (X, t) are
uniformly bounded with respect to X and t.
Sketch of the proof. First of all, we notice that
lj ∂ k lj
hk (x, t) = (−i)k h (x, t).
∂Elk 0
As in Lemma 10, we obtain

1 iEl ξ l2
0 (x, t) = e
hl1 B0 (X, El ) 1 + εl1 (X, t) , 0 (x, t) = √ e
iEl ξ
hl2 h00 (X, t),
t
where εl1 (X, t) = o(1) for t → ∞, hl2
00 (X, t) is uniformly bounded. These relations
may be differentiated with respect to El , and the estimates will be still valid.
COROLLARY 6. Since
x C ξ 1 ln t
X= = + = + O ,
2t 2 2t 2|El |2 t
the asymptotic equality

lj lj ln t
hkm (X, t) = hkm (t) 1 + ON
t
holds.
LEMMA 14. For coefficients (5.9), the relations
X
N−k−1 X
m X
s
−1/2 (l)
ks (x, t) = t
Alr lrpq lr
ωkm (t) Hms (t)Jp+q (ξ ),
m=0 p=0 q=0
X
N−k−1
(l)
X
m X
s
lr
Bks (x, t) = ωkm (t) Glrpq lr
ms (t)Jp+q (ξ ),
m=0 p=0 q=0
where
Z ∞
Jplr (ξ ) = τ p exp[i(Er + El )τ ] dτ,
ξ
lrpq
Hms (t) = hl1
mp (t)hsq (t) + hmp (t)hsq (t),
r2 l2 r1
ms (t) = hmp (t)hsq (t) + hmp (t)hsq (t)

Glrpq l1 r1 l2 r2
(l) lrpq lrpq

are valid with the functions ωkm (t), Hms (t), and Gms (t), that are uniformly
bounded with respect to t.
The system of Equations (5.8) reduces to the following system
X X
p N−1
Xkl + Ys = 0,
lr r
Bks l = 1, 2, . . . , p,
r=1 s=0
X X
p N−1
Xs + Ykl = bkl ,
lr r
Bks k = 0, 1, . . . , N − 1.
r=1 s=0
p
Let us introduce the block matrix (ξ, t) = kB lr (ξ, t)klr=1 . Then the system of
these equations takes the form
X + Y = 0,
X + Y = b.
Therefore
1
Y = (I + )−1 b + (I − )−1 b .
2
Equation (5.7) leads to the representation
∂ det[I + (ξ, t)]
v(x, t) = 2i ln + o(1), ξ = x − Ct, t → ∞,
∂ξ det[I + (ξ, t)]
for the solution of the problem (3), (5). Hence Equation (7) is valid and Theorem 2
is proved.
Now let us shortly consider Theorem 3. In this case, the entries of the matrix
(x, t) have the form
X
N−k−1
kj (ξ, t) = ωkm (t)wmj (ξ, t),
m=0
X
m X
j

wmj (ξ, t) = h1mp (t)h1j q (t) + 1 + O t −1 Ip+q (ξ ),
p=0 q=0
Z ∞
In (ξ ) = τ n e−2aτ dτ,
ξ
where
−1
ωkm (t) = −ih0 0(k + m + 3/2) k!m!(4a)2(k+m+1) t k+m+3/2 [1 + ψkm (t)]
with some constant h0 , the gamma-function 0(k + m + 3/2) and an estimate for
2
k + m2
ψkm (t) = O √ .
t
The determinant 1(ξ, t) = det[I + (ξ, t)] can be written in a standard way
X
N X
1(ξ, t) = Dk (ξ, t), Dk (ξ, t) = det i1 ...ik (ξ, t).
k=1 i1 <i2 <···<ik
If n 6 [ N+1
2
], then the determinant of the matrix kwmj (x, t)k of n + 1 order can be
represented as follows

det wmj (x, t)
1 1
h00 0 . . . 0 I0 I1 . . . In h00 h10 . . . hn0
1 1

h 1 h 1 . . . 0 I1 I2 . . . In+1 0 h111 . . . h1n1
10
= . .. .
11
.. .. .. .. .. .. ..
. . . . . .. . .
h1 h1 . . . h1 In In+1 . . . I2n 0 0 . . . h 1
n0 n1 nn nn
Y n

h1kk det Im+j (ξ ) = B0 (X, ia) det Im+j (ξ ) .
2 2n+2
=
k=0
By using the technique from [12] we find
X
N

1(ξ, t) = 1 + i n Pn (X, t)t −n(n+1/2) e−2naξ 1 + δn (t) , (5.10)
n=1
where

(−h0 )n B02n (X, ia)0 (n) 01(n) N +1
Pn (X, t) = Q , n6 ,
(4a)2n2 (2a)n2 ( n−1
k=0 k!)
2 2

N +1
Pn (X, t) = O(1), n >
2
with some δn (t) = O(n2 /t). This expansion leads to Equations (10), (11) with

1 |h0 | 1 0 (n) 01(n)
αn0 = ln , (5.11)
2a [(n − 1)!]2 (2a)6n−3 0 (n−1) 01(n−1)
where 0 (n) and 01(n) are the determinants with the gamma-function entries 0(k +
j + 3/2) and 0(k + j + 1). Theorem 3 is proved.
Below we give a sketch of the proof of Theorem 4. In this case

A(ξ, t) B(ξ, t)
(ξ, t) = ,
−B(ξ, t) −A(ξ, t)
where
X
N−k−1
Akj (ξ, t) = 1
ωkm (t)wmj (ξ, t),
m=0
X
N−k−1
Bkj (ξ, t) = 2
ωkm (t)wmj (ξ, t),
m=0
X
m X
j

1
wmj (ξ, t) = h1mp (t)h1j q (t) 1 + O t −1 Jp+q (ξ ),
p=0 q=0
X
m X
j

2
wmj (ξ, t) = h1mp (t)h1j q (t) 1 + O t −1 Ip+q (ξ ),
p=0 q=0
Z ∞ Z ∞
Jn (ξ ) = n 2iEτ
τ e dτ, In (ξ ) = τ n e−2 Im Eτ dτ,
ξ ξ
i n h0 eid0 t c1n+1 0(n + 3/2)

ωkm (t) = ×
(k!m!)t n+3/2 (1 − id1 )n+3/2
√
× [1 + ψn (t)], n = k + m, |ψn (t)| 6 Cn2 / t.
Using the same scheme as in [12], we notice that

A B ω 0 h− 0 J I h+ 0
1 (ξ, t) =
(k,l) = = ,
−B −A 0 −ω 0 h̄− I J 0 h̄+
where
 
ω00 ω01 ··· ω0k−1
 ω10 ω11 ··· ω1k−1 
ω=
 ... .. .. .. ,

. . .
ωk−10 ωk−11 ··· ωk−1k−1
 1 
h00 0 ··· 0
 h1 h111 ··· 
 0 
h− =  .10 .. .. .. ,
 .. . . . 
h1k−10 h1k−11 · · · h1k−1k−1
   
J0 ... Jk−1 I0 ... Ik−1
J =  ... ..
.
..  ,
. I =  ... ..
.
..  ,
.
Jk−1 . . . J2k−2 Ik−1 ... I2k−2
 1 
h00 h110 ··· h1k−10
 0 h111 ··· h1k−11 
 
h+ =  . .. .. .. .
 .. . . . 
0 0 ··· h1k−1k−1
This allows us to obtain an expansion analogous to (5.10), which leads to Equa-
tions (12)–(14) with

1 Re E 2n−1
βn =
0
(2n − 1) arctan + arg h0 c1 (1 − id1 ) 1/2−2n
, (5.12)
2 Re E Im E

1 |h0 ||c1 |2n−1 01(n) | Re E| 2n−1
γn0 = ln . (5.13)
2 Im E [(n − 1)!2 ]|1 − id1 |2n−1/2 01(n−1) 2|E| Im E
Theorem 4 is proved.
Acknowledgements
V.K. thanks the Fakultät und Institut für Mathematik, Ruhr Universität Bochum for
kind hospitality and the DFG for financial support.
References
1. Novikov, S. P., Manakov, S. V., Pitaevsky, L. P. and Zakharov, V. E.: Theory of Solitons. The
Inverse Scattering Method, Nauka, Moscow, 1980, 319 pp. English transl. Plenum, New York,
1984.
2. Faddeev, L. D. and Takhtadjan, L. A.: Hamiltonian Methods in the Theory of Solitons, Nauka,
Moscow, 1986. English transl. Springer-Verlag, 1987.
3. Schuur, P. C.: Asymptotic analysis of soliton problems, Lect. Notes in Math. 1232, Springer-
Verlag, 1986.
4. Its, A. R.: Method of the isomonodromic deformation in the theory of completely integrable
nonlinear evolution systems, Doctoral dissertation, Leningrad University, 1986.
5. Novokshenov, V. Yu.: Asymptotics as t → ∞ of the solution of the Cauchy problem for the
nonlinear Schrödinger equation, Sov. Math. Dokl. 21(2) (1980), v. 251, 529–533.
6. Kotlyarov, V. P.: Asymptotic analysis of the Marchenko integral equation and soliton
asymptotic of a solution of the nonlinear Schrödinger equation, Physica D 87(1–4) (1995),
176–185.
7. Shabat, A. B.: On the Korteweg–de Vries equation, Sov. Math. Dokl. 14 (1973), 1266–1269.
8. Khruslov, E. Ya.: Asymptotics of the solution of the Cauchy problem for the Korteweg–de
Vries equation with the step-like initial data, Mat. Sbornik 99(2) (1976), 261–281.
9. Fokas, A. S. and Zakharov, V. E. (eds.): Important Developments in Soliton Theory, Springer-
Verlag, 1993.
10. Fokas, A. S. and Its, A. R.: An initial-boundary value problem for the sine-Gordon equation in
laboratory coordinates, Teoreticheskaya i Matematicheskaya Fizika 92(3) (1992), 387–403.
11. Fokas, A. S. and Its, A. R.: The linearization of the initial-boundary value problem of the
nonlinear Schrödinger equation, SIAM J. Math. Anal. 27(3) (1996), 738–764.
12. Kotlyarov, V. P.: Asymptotic solitons of the sine-Gordon equation, Theor. and Math. Physics
80(1) (1989), 15–28.
13. Kotlyarov, V. P.: Influence of double continuous spectrum of the Dirac operator on the nonlinear
Schrödinger equation asymptotic solitons, Matem. Zametki 49(2) (1991), 84–94. (Math. Notes
(1991) (1–2), 172–180.)
14. Reed, M. and Simon, B.: Methods of Modern Mathematical Physics. 1. Functional Analysis,
Academic Press, NY, 1972.
53
On the Davey–Stewartson and Ishimori Systems
NAKAO HAYASHI
Department of Applied Mathematics, Science University of Tokyo, 1-3, Kagurazaka, Shinjukuku,
Tokyo 162, Japan; e-mail: nhayashi@rs.kagu.sut.ac.jp
PAVEL I. NAUMKIN
Instituto de Física y Matemáticas, Universidad Michoacana AP 2-82, CP 58040, Morelia,
Michoacan, Mexico; e-mail: naumkin@ifm1.ifm.umich.mx
(Received: 30 September 1998, accepted: 16 March 1999)

Abstract. We study the initial value problem for the two-dimensional nonlinear nonlocal Schrödin-
ger equations
iut + 1u = N (v), (t, x, y) ∈ R3 , u(0, x, y) = u0 (x, y), (x, y) ∈ R2 , (A)
where the Laplacian 1 = ∂x2 + ∂y2 , the solution u is a complex valued function, the nonlinear term
N = N1 +N2 consists of the local nonlinear part N1 (v) which is cubic with respect to the vector v =
(u, ux , uy , u, ux , uy ) in the neighborhood of the origin, and the nonlocal nonlinear part N2 (v) =
Rx
(v, ∂ −1 Kx (v)) + (v, ∂y−1 Ky (v)), where (·, ·) denotes the inner product, ∂x−1 = −∞ dx 0 , ∂y−1 =
Ry x 0
−∞ dx and the vectors Kx ∈ (C (C ; C)) and Ky ∈ (C (C ; C)) are quadratic with respect
4 6 6 4 6 6
(2) (4)
to the vector v in the neighborhood of the origin. We assume that the components Kx = Kx ≡
(3) (6)
0, Ky = Ky ≡ 0. In particular, Equation (A) includes two physical examples appearing in fluid
dynamics. The elliptic–hyperbolic Davey–Stewartson system can be reduced to Equation (A) with
(1) (1)
N1 = |u|2 u, Kx = ∂y (|u|2 ), Ky = ∂x (|u|2 ), and all the rest components of the vectors Kx
and Ky are equal to zero. The elliptic–hyperbolic Ishimori system is involved in Equation (A), when
N1 = (1 + |u|2 )−1 u(∇u)2 , and Kx = −Ky = (1 + |u|2 )−2 (ux uy − ux uy ). Our purpose in this
(3) (2)
paper is to prove the local existence in time of small solutions to the Cauchy problem (A) in the usual
Sobolev space, and the global-in-time existence of small solutions to the Cauchy problem (A) in the
weighted Sobolev space under some conditions on the complex conjugate structure of the nonlinear
terms, namely if N (eiθ v) = eiθ N (v) for all θ ∈ R.
Mathematics Subject Classification (1991): 35Q55.
Key words: Davey–Stewartson system, Ishimori system, elliptic–hyperbolic case, nonlocal nonlinear
Schrödinger equation.
1. Introduction
We study the initial value problem for the two-dimensional nonlinear nonlocal
Schrödinger equations

iut + 1u = N (v), (x, y) ∈ R2 , t ∈ R,
(1.1)
u(0, x, y) = u0 (x, y), (x, y) ∈ R2 ,
54 N. HAYASHI AND P. I. NAUMKIN
where the solution u is a complex valued function, the Laplacian is 1 = ∂x2 + ∂y2 ,
and the nonlinear term N (v) = N1 (v) + N2 (v) consists of the local nonlinear part
N1 (v), which is cubic with respect to the vector v = (u, ux , uy , ū, ūx , ūy ) in the
neighborhood of the origin, and the nonlocal nonlinear part N2 (v) =R (v, ∂x−1 Kx (v))
+ (v, ∂y−1 Ky (v)), where (·, ·) denotes the inner product, ∂x−1 = −∞ dx 0 , ∂y−1 =
x
Ry 0
−∞ dy and the vectors Kx and Ky are quadratic with respect to the vector v in the
neighborhood of the origin. We assume that N1 (z) ∈ C4 (C6 ; C), N1 (z) = O(|z|3 )
as z → 0, Kx (z), Ky (z) ∈ (C4 (C6 ; C))6 , Kx (z) = O(|z|2), Ky (z) = O(|z|2 )
as z → 0. We also assume that the components Kx(2) ≡ Kx(4) ≡ 0, and Ky(3) ≡
Kx(6) ≡ 0.
We show how to translate the Davey–Stewartson and Ishimori systems to the
nonlinear nonlocal Schrödinger Equation (1.1) and we include some historical
comments and remarks about previous results for these systems. The Davey–Ste-
wartson (D-S) system is written as

 iut + αuxx + uyy = a|u|2 u + buωx , (x, y) ∈ R2 , t ∈ R,
ωxx + βωyy = ∂x |u|2 , (x, y) ∈ R2 , t ∈ R, (1.2)

u(0, x, y) = u0 (x, y), (x, y) ∈ R2 ,
where α, β ∈ R, a, b ∈ C. This system was derived by Davey and Stewartson
[11], Benney and Roskes [6] and Djordjevic and Redekopp [12] for describing
the evolution of weakly nonlinear water waves that travel predominantly in one
direction, and where the wave amplitude is modulated slowly in the horizontal
direction. Independently, Ablowitz and Haberman [1] and Cornille [10] obtained
a particular form of system (1.2) as an example of a completely integrable system
generalizing the one-dimensional Schrödinger eguation. Djordjevic and Redekopp
in [12] have shown that the parameter β can become negative when the capillarity
effects are important.
In the literature on the inverse scattering transform method the Davey–Stewart-
son system (1.2) with α = 1, β = −1, a = −1, b = 2, is referred to as the
DSI system. And if α = −1, β = 1, a = 2, b = −1, or α = −1, β = 1, a =
−2, b = 1, system (1.2) is called the DSII focusing system or the DSII defocusing
system, respectively. For these cases the existence of soliton-like solutions or so-
called lump solutions were obtained in [4] and [14], and the existence of solutions
to the Cauchy problem was proved via the inverse scattering transform method (see
[2, 3, 10, 15, 33]). Using the inverse scattering transform method, Sung [33] proved
the global existence and uniqueness for the Cauchy problem to DSII systems with
any (large) initial data û0 ∈ L1 (R2 ) ∩ L∞ (R2 ) for the defocusing case, and with
small initial data in the same space for the focusing case. Moreover, he proved some
regularity results and qualitative properties; for instance, that the solution u(t, x, y)
disperses to zero in L∞ (R2 ) as t → ∞. Beals and Coifman [7] proved the global
well-posedness of the Cauchy problem for the DSII defocusing system when the
initial data are in the Schwartz class. The same result was obtained by Fokas and
Sung [15] for the DSI system. However, all the cases of the Davey–Stewartson
DAVEY–STEWARTSON AND ISHIMORI SYSTEMS 55
system (1.2) to which the inverse scattering transform method can be applied are
very exceptional. Note also that the inverse scattering transform method is rather
complicated. In comparison with other functional analytical methods it leads to
more restrictions on the initial data.
Ghidaglia and Saut [16] classified the Davey–Stewartson system (1.2) as elliptic–
elliptic, elliptic–hyperbolic, hyperbolic–elliptic and hyperbolic–hyperbolic accord-
ing to the signs of the coefficients (α, β) as follows: (+, +), (+, −), (−, +)
and (−, −) respectively. In [16] they proved some results on the local and global
existence of solutions in the usual Sobolev spaces L2 (R2 ), H1 (R2 ) and H2 (R2 ) for
the more simple elliptic–elliptic and hyperbolic–elliptic cases. Also, they obtained
a blow-up effect for the solutions of the elliptic–elliptic Davey–Stewartson system.
Some results on the asymptotic behavior in time of solutions for the elliptic–elliptic
case were found by Constantin in [9]. As far as we know, the first results on the
local existence and uniqueness of solutions to the Cauchy problem for the elliptic–
hyperbolic and hyperbolic–hyperbolic Davey–Stewartson systems with sufficiently
small initial data were obtained via the functional analytical methods by Linares
and Ponce in [30]. They used a sharp version of the smoothing effect obtained by
Kato [27] for the evolution groups {eit ∂x ∂y }∞−∞ and {e
it 1 ∞
}−∞ .
Now let us reduce the Davey–Stewartson systems in the elliptic–hyperbolic case
to Equation (1.1). Without a loss of generality we now take β = −1 in system
(1.2). Let us assume that the function ω satisfies the following radiation condition
ω(x, y, t) → 0 as x + y → −∞ and x − y → −∞ in order to solve the second
equation of system (1.2). This condition guarantees the existence of the inverse
operator (∂x2 − ∂y2 )−1 in L1 (R2 ). Thus, after a rotation in the xy-plane: x = √12 (x 0 +
y 0 ), y = √12 (x 0 − y 0 ) the elliptic–hyperbolic Davey–Stewartson system (we have
chosen α = 1) takes the form (the primes we omit):
iut + 1u = ã|u|2 u + b̃u∂x−1 ∂y |u|2 + b̃u∂y−1 ∂x |u|2 , (1.3)

Rx Ry
where ∂x−1 = −∞ dx 0 , ∂y−1 = −∞ dy 0 , ã = a + b2 , b̃ = b4 are arbitrary com-
plex constants. Therefore, as we already mentioned above, the elliptic–hyperbolic
Davey–Stewartson system is a particular case of Equation (1.1) if we choose N1 =
ã|u|2 u, Kx(1) = b̃∂y (|u|2 ), Ky(1) = b̃∂x (|u|2 ) and all the rest components of
Kx , Ky to be equal to zero identically.
We now turn to the Ishimori system

 Et = E × (Exx + αEyy ) + a(ωx Ey + ωy Ex ), (x, y) ∈ R2 , t ∈ R,
ωxx + βωyy = bE · (Ex × Ey ), (x, y) ∈ R2 , t ∈ R, (1.4)

E(0, x, y) = E0 (x, y), (x, y) ∈ R2 ,
where the coefficients are α = −1, β = 1, b = 2, a ∈ R or α = 1, β = −1, b =

−2, a ∈ R, the unknown function E(t, x, y) is a vector in R3 such that |E|2 = 1,
and E → (0, 0, 1) as |x| + |y| → ∞; here the operation × denotes the outer
product in R3 .
Ishimori [26] proposed system (1.4) as a two-dimensional generalization of

the Heisenberg equation in ferromagnetism. When α = 1, a = 0 the Ishimori
system (1.4) can be reduced to the two-dimensional Heisenberg equation (or to
the following nonlinear Schrödinger equation with derivatives in the nonlinearity
iut + 1u = 1+|u| 2ū 2
2 (∇u) (see [5] and also Equations (1.5), (1.6) below)). In the
case of a = 1, system (1.4) was studied formally in [6] and [29] by virtue of the
inverse scattering transform method.
Via the stereographic projection we translate the Ishimori system (1.4) to a
system of two scalar equations. Let us put
−1
E = E (1) , E (2) , E (3) = 1 + |u|2 u + ū, −i(u − ū), 1 − |u|2 ,
so that u = (E (1) +iE (2) )/(1+E (3)) is a single complex valued unknown function.
Thus the restriction |E| = 1 is fulfilled automatically, and the boundary condition
E → (0, 0, 1) as |x| + |y| → ∞ leads to a natural decay property of the solution
u at infinity: u → 0 as |x| + |y| → ∞. Thus we get

 iut + uxx + αuyy


 2ū
= (ux )2 + α(uy )2 + ia(ux ωy + uy ωx ), (x, y) ∈ R2 , t ∈ R,
1 + |u| 2 (1.5)

 2 −2
 ωxx + βωyy = 2ib(1 + |u| ) (ux ūy − uy ūx ),
 (x, y) ∈ R , t ∈ R,
2
u(0, x, y) = u0 (x, y), (x, y) ∈ R2 .

Similarly to the classification of the Davey–Stewartson system (1.2), we classify
the Ishimori system (1.5) as elliptic–elliptic, elliptic–hyperbolic, hyperbolic–elliptic
or hyperbolic–hyperbolic cases according to the signs of the coefficients (α, β) :
(+, +), (+, −), (−, +) and (−, −) respectively. The local well-posedness and the
global existence of small solutions to the Cauchy problem for the Ishimori system
(1.5) in the more simple hyperbolic–elliptic case was studied by Soyeur in [31] (he
choose α = −1, β = 1, a = 2, b = 2. It is clear that his method can also be
applied to the Ishimori system (1.5) in the elliptic–elliptic case). Using the gauge
equivalence between the integrable Ishimori system and focusing the DSII system,
Sung [34] proved the global existence and uniqueness of solutions to the Cauchy
problem for the integrable Ishimori system with small initial data. He also obtained
some regularity results, in particular the existence of solutions in the Schwartz
class.
As in the case of the elliptic–hyperbolic Davey–Stewartson system via a rotation
in the xy-plane x = √12 (x 0 + y 0 ), y = √12 (x 0 − y 0 ), we can write the elliptic–
hyperbolic Ishimori system (1.5) (with α = 1, β = −1) in the following manner:
2ū
iut + 1u = (ux )2 + (uy )2 + uy ∂x−1 Kx(3) + ux ∂y−1 Ky(2), (1.6)
1 + |u|2
where Kx(3) = −Ky(2) = ã(1 + |u|2 )−2 (ux ūy − ūx uy ) and ã = 2ab is an arbitrary
real constant. Thus the elliptic–hyperbolic Ishimori system is a particular case of
Equation (1.1) with N1 = 2(1 + |u|2 )−1 ū(∇u)2 , and Kx(3) = −Ky(2) = a(1 +
|u|2 )−2 (ux ūy − ūx uy ) and when all the other components of the vectors Kx and
Ky are equal to zero.
Note that the standard Lp –Lq time-decay estimates with the classical energy
method are useful for the nonlinear Schrödinger equation; however, they are not
applicable directly to Equations (1.3) and (1.6), because of the presence of deriv-
atives and nonlocal terms in the nonlinearities. There are few papers devoted to
Equations (1.3) and (1.6). In [24] and [25] the global existence of solutions to the
Cauchy problem for these equations with small analytic initial data was proved.
The local and global existence of small solutions for Equation (1.3) were also
shown in [17] and [21] in suitable weighted Sobolev spaces and in [8] and [22]
the usual Sobolev spaces were applied for proving the local existence of small
solutions for (1.3). Later the smallness condition was removed in [19].
The purpose of this paper is to study the nonlinear nonlocal Schrödinger equa-
tion (1.1), involving the elliptic–hyperbolic Davey–Stewartson (1.3) and Ishimori
(1.6) systems in the usual weighted Sobolev spaces. Our results of Theorems 1.1,
1.2, 1.4 improve the previous results on these systems [8, 20, 22 – 24, 30]. As far
as we know, the result of Theorem 1.3 is the first for the problem of the global
existence and the asymptotic behavior for large time of small solutions of the
elliptic–hyperbolic Ishimori system. Section 4 is devoted to the proof of local-
in-time existence of small solutions to the Cauchy problem for (1.1). Let us denote
the usual Sobolev space as follows: Hm,0 = {φ ∈ L2 ; k(1 − ∂x2 − ∂y2 )m/2 φk < ∞},
here and below k·k is the norm of the Lebesgue space L2 . For simplicity we denote
the norm of the Sobolev space Hm,0 as follows: k · kHm,0 = k · km,0 . We prove the
following result.
THEOREM 1.1. Let the initial data u0 ∈ H4,0 and the norm ku0 k2,0 be sufficiently
small. Then there exists a time T > 0 such that there exists a unique solution
u ∈ C([0, T ]; H3,0 ) ∩ L∞ (0, T ; H4,0 ) of the Cauchy problem (1.1). Moreover, if
the norm ε = ku0 k4,0 of the initial data is sufficiently small then there exists a time
T > 1 and a unique solution u ∈ C([0, T ]; H3,0 ) ∩ L∞ (0, T ; H4,0 ) to the Cauchy
problem (1.1) such that kuk4,0 6 Cε for all t ∈ [0, T ].
If we assume additional constraints on the nonlinear terms we can relax the
regularity condition on the initial data.
THEOREM 1.2. Let the local nonlinear part N1 and Kj(l) , j = x, y, l = 2, 3, 5, 6,

depend on u, ū only, and Kj(1), Kj(4) , j = x, y, are linear with respect to ∇u and
∇ ū. We also assume that the initial data u0 ∈ H2,0 and have sufficiently small
norm ku0 k. Then there exists a time T > 0 such that there exists a unique solution
u ∈ C([0, T ]; H1,0 ) ∩ L∞ (0, T ; H2,0 ) of the Cauchy problem (1.1). Moreover, if
the norm ε = ku0 k2,0 is sufficiently small then there exists a time T > 1 and a
unique solution u ∈ C([0, T ]; H1,0 ) ∩ L∞ (0, T ; H2,0 ) such that kuk2,0 6 Cε for
all t ∈ [0, T ].
REMARK 1.1. Theorems 1.1 and 1.2 give us the local existence in time of so-
lutions to the elliptic–hyperbolic Davey–Stewartson and Ishimori systems in H2,0
and H4,0, respectively. These results are improvements of the previous works [22]
and [20]. Note that an additional smallness condition on the initial data allows us
to choose the existence time T > 1.
In Sections 4 and 5 we study the global-in-time existence and asymptotic be-

havior of solutions to the Cauchy problem (1.1), when the nonlinear terms satisfy
the following condition such that N (eiθ v) = eiθ N (v) for all θ ∈ R; therefore,
we assume that Kj(l) (v) = 0, j = x, y, l = 4, 5, 6. As we already mentioned
above, the local nonlinear part N1 (v) is cubic with respect to the vector v =
(u, ux , uy , ū, ūx , ūy ) in the neighborhood of the origin, and the components Kj(l) ≡
Kj(l) (v, v̄), j = x, y, l = 1, 2, 3, can be quadratic with respect to v. We may write
them as follows
Kj(1) = λ(1) (2) (3) (4) (5) (6)

j |u| + iλj uūx + iλj ūux + iλj uūy + iλj ūuy + λj |ux | +
2 2
+ λj(7) |uy |2 + λj(8)ux ūy + λ(9) (1)

j ūx uy + Bj ,
Kx(3) = iµ(1) |u|2 + µ(2) uūx + µ(3) ūux + µ(4) uūy + µ(5) ūuy + iµ(6) |ux |2 +
+ iµ(7) |uy |2 + iµ(8) ux ūy + iµ(9) ūx uy + Bx(3) , and
Ky(2) = iν (1) |u|2 + ν (2) uūx + ν (3) ūux + ν (4) uūy + ν (5) ūuy + iν (6) |ux |2 +
+ iν (7) |uy |2 + iν (8) ux ūy + iν (9) ūx uy + By(2) , where
Bj(l) = O(|v|3 ), l = 1, 2, 3, j = x, y.
We assume that the coefficients of these representations satisfy the following con-
ditions
λ(1)
j , λj(2) − λj(3) , λj(4) − λj(5), λj(6) , λj(7) , λj(8) + λj(9) ∈ R,
µ(1) , µ(2) − µ(3) , µ(4) − µ(5) , µ(6) , µ(7) , µ(8) + µ(9) ∈ R,
ν (1) , ν (2) − ν (3) , ν (4) − ν (5) , ν (6) , ν (7) , ν (8) + ν (9) ∈ R, (1.7)
where j = x, y.
REMARK 1.2. In the case of the elliptic–hyperbolic Davey–Stewartson system

we have λ(4)x = λx = −i b̃, λy = λy = −i b̃ and the rest coefficients λ, µ, ν
(5) (2) (3)
are zero. We get the elliptic–hyperbolic Ishimori system when ν (8) = −ν (9) =
−i ã, µ(8) = −µ(9) = −i ã, and the other coefficients λ, µ, ν are equal to zero.
Thus the conditions (1.7) are fulfilled for these examples.
In Section 4 we prove the global existence of solutions to the Cauchy problem

(1.1) using the weighted Sobolev spaces

Hm,l,p = φ ∈ L2 ; (1 − ∂x2 − ∂y2 )m/2 (1 + x 2 + y 2 )l/2 φ p < ∞ .
We denote the norm of the weighted Sobolev space Hm,l,p by k · kHm,l,p = k · km,l,p .
p p q
For simplicity we write Hm,l = Hm,l,2 , Lj = Lp (Rj ), Lx Ly = Lp (Rx ; Lq (Ry )),
Hm,l
j = Hm,l (Rj ), and we take into account the ordering of the norms, denoting
kφkY2 X1 = k kφkX1 kY2 , where X1 and Y2 are some Banach spaces. For exam-
ple: kφkCL2 = k kφ(t)kL2 kC = supt kφ(t)kL2 , etc. We denote by U(t) the linear
Schrödinger evolution group. We prove the following result.
In the following four theorems we always assume that condition (1.9) is fulfilled
and Kj(l) = 0, j = x, y, l = 4, 5, 6.
THEOREM 1.3. We assume that the initial data u0 ∈ H4,1 and the norm ku0 k4,1 =
ε is sufficiently small. Then there exists a unique solution u ∈ C(R; L2 ) of the
Cauchy problem (1.1) such that
p
(1 + |t|)−γ kU(−t)u(t)k4,1 + 1 + |t|(k(1 − 1)u(t)kL∞ 2+
x Ly
+ k(1 − 1)ux (t)kL∞ 2 + k(1 − 1)u(t)kL∞ L2 + k(1 − 1)uy (t)kL∞ L2 )

x Ly y x y x
6 Cε
for all t ∈ R, where γ > 0 is sufficiently small and C is some positive constant.
For the case of Theorem 1.2 we have the corresponding global existence result.
THEOREM 1.4. We assume that the local part N1 and Kj(l) , j = x, y, l =

2, 3, 5, 6, depend on u, ū only, and Kj(1) , j = x, y, are linear with respect to ∇u
and ∇ ū. We also assume that the initial data u0 ∈ H2,1 and the norm ku0 k2,1 = ε is
sufficiently small. Then there exists a unique solution u ∈ C(R, L2) of the Cauchy
problem (1.1) such that
p
(1 + |t|)−γ kU(−t)u(t)k2,1 + 1 + |t| ku(t)kL∞ 2 + kux (t)kL∞ L2 +
x Ly

x y
+ ku(t)kL∞ 2 + kuy (t)kL∞ L2 6 Cε

y Lx y x
for all t ∈ R, where γ > 0 is sufficiently small and C > 0 is some constant.
If the initial data decay more rapidly at infinity we will obtain in Section 5 the
decay estimate in the uniform norm and the large time asymptotic behavior of small
solutions to the Cauchy problem (1.1).
THEOREM 1.5. We assume that the initial data u0 ∈ H4,1 ∩ H3,2 and the norm
ku0 k4,1 + ku0 k3,2 = ε is sufficiently small. Then there exists a unique solution
u ∈ C(R, L2) of the Cauchy problem (1.1) such that
(1 + |t|)−γ (ku(t)k4,1 + kU(−t)u(t)k3,2 ) + (1 + |t|)k(1 − 1)u(t)k∞ 6 Cε
for all t ∈ R, where γ > 0 is sufficiently small and C is some positive constant.
Moreover, there exists a unique final state W + ∈ L2 ∩ L∞ and the real-valued
phase function Q+ ∈ L∞ such that the solution u of the Cauchy problem (1.1) has
the following asymptotics:
(1 − 1)u(t) = t −1 W + (χ)exp(iχ 2 t + iQ+ (χ) log t) + O(t −1−δ ), (1.8)
as t → ∞ uniformly in x ∈ R2 , where χ = ( 2tx , 2ty ), δ ∈ (0, 1/4) is some constant.
The next result corresponds to the case of Theorem 1.2.
THEOREM 1.6. We assume that the local part N1 and Kj(l) , j = x, y, l = 2, 3
depend on u, ū only, and Kj(1) , j = x, y, are linear with respect to ∇u and ∇ ū. We
also assume that the initial data u0 ∈ H2,1 ∩H1,2 and the norm ku0 k2,1 +ku0 k1,2 =
ε is sufficiently small. Then there exists a unique solution u ∈ C(R, L2 ) of the
Cauchy problem (1.1) such that
(1 + |t|)−γ (ku(t)k2,1 + kU(−t)u(t)k1,2 ) + (1 + |t|)ku(t)k∞ 6 Cε
for all t ∈ R, where γ > 0 is sufficiently small and C > 0 is some constant.
Moreover, there exists a unique final state u+ ∈ L2 ∩ L∞ and the real-valued
phase function Q+ ∈ L∞ such that the solution u of the Cauchy problem (1.1) has
the asymptotics
u(t) = t −1 u+ (χ) exp(iχ 2 t + iQ+ (χ) log t) + O(t −1−δ ),
as t → ∞ uniformly in x ∈ R2 , where χ = ( 2tx , 2ty ), δ > 0 is some constant
satisfying 0 < δ < 1/4.
REMARK 1.3. Theorems 1.3 and 1.4 show the global-in-time existence of small
solutions to the elliptic–hyperbolic Ishimori and Davey–Stewartson systems, re-
spectively. For the cases of the Ishimori and Davey–Stewartson systems we have
Q+ ≡ 0, therefore Theorems 1.5 and 1.6 show the existence of the usual scattering
states for these cases.
We organize the rest of the paper as follows. First we give the notations which
we use in this paper. Then in Section 2 we prepare some preliminary estimates.
In Lemma 2.1 we estimate commutators of the fractional derivatives of order α ∈
(0, 1). In Lemma 2.2 we obtain a smoothing property of the linear part of the
Schrödinger equations (1.1). Finally in Lemma 2.3 we estimate different terms
appearing in the nonlinearity. In Section 3 we prove Theorems 1.1 and 1.2 which
state the local existence of solutions to the Cauchy problem (1.1). Section 4 is
devoted to the proof of Theorems 1.3 and 1.4. And in Section 5 we prove Theorems
1.5 and 1.6.
R xj
Notations. We denote ∂j = ∂x∂ j , ∂j−1 = −∞ dxj0 , x1 ≡ x, x2 ≡ y, ∂x−1 =
∂1−1 , ∂y−1 = ∂2−1 . Let Fj φ or φ̂ be the Fourier transform of φ(xj ), namely φ̂(ξj ) =
R −iξ x
√1
2π R
e j j φ(xj ) dxj . We denote by Fj−1 φ or φ̌ the inverse Fourier transform
R
of the function φ(ξj ), defined by the formula φ̌(xj ) = √1
2π R
eiξj xj φ(ξj ) dξj . In
what follows we also use the following relation |∂j | = Fj−1 |ξj |Fj = −Hj ∂j . The
Hilbert transformation with respect to the variable x is defined as follows:
Z
1 φ(z, y) ξx
H1 φ(x, y) ≡ Hx φ(x, y) = PV dz = −iFx−1 Fx φ,
π R x −z |ξx |
where PV means the principal value of the singular integral. The operator H2 is
defined in the same way. We widely use the fact that the Hilbert transformations
Hj are bounded operators from L2j to L2j , j = x, y. The fractional derivative
|∂x |α , α ∈ (0, 1) is equal to
Z
dz
|∂x |α φ = Fx−1 |ξx |α Fx φ = C (φ(x + z, y) − φ(x, y)) 1+α
R |z|
and similarly we have
Z
dz
|∂x | Hx φ =
α
−iFx−1 signξx |ξx |α Fx φ =C (φ(x + z, y) − φ(x, y)) ,
R z|z|α
with some constant C, see [32]. The derivatives |∂y |α and |∂y |α Hy are defined
analogously. Let J = (J1 , J2 ) = (Jx , Jy ), where Jj = xj + 2it∂j .
2. Linear Smoothing Effect

The aim of this section is to obtain the smoothing effect for solutions to the Cauchy
problem for the linear Schrödinger equations

iut + 1u = f, (x, y) ∈ R2 , t ∈ R,
(2.1)
u(0, x, y) = u0 (x, y), (x, y) ∈ R2 ,
where 1 = ∂x2 + ∂y2 and the function f (t, x, y) is a force. Below in Sections 3
and 4 we will consider the nonlinearity instead of f (t, x, y). Smoothing effects for
solutions to the Cauchy problem (2.1) were studied by many authors (see [13] and
[28] and references therein). In our key Lemma 2.2 of this section we will obtain a
simple and explicit modification of smoothing effects of Doi [13].
In the next lemma we prove that if φ is a sufficiently smooth function, then the
commutators [|∂j |α , φ], and [|∂j |α Hj , φ], j = x, y, are continuous operators from
L2j and L2j .
LEMMA 2.1. The following inequalities

k[|∂x |α , φ(x, y)]ψ(x, y, η)kL2x Lqη Lpy 6 C kφkL∞ s + kφx kL∞ Ls kψk 2 q r
x Ly x y Lx Lη Ly
and

k[|∂x |α Hx , φ(x, y)]ψ(x, y, η)kL2x Lqη Lpy 6 C kφkL∞ s + kφx kL∞ Ls kψk 2 q r
x Ly x y Lx Lη Ly
are valid for any 1 6 p, q 6 ∞, where 1r + 1s = p1 , provided that the right-hand

sides are bounded.
Proof. By virtue of the Hölder inequality we get
k[|∂x |α , φ(x, y)]ψ(x, y, η)kL2x Lqη Lpy
Z
dz

= C ψ(x + z, y, η)(φ(x + z, y) − φ(x, y)) 1+α
R |z| 2 q p
Lx Lη Ly
Z Z
z
dz

6 C |ψ(x + z, y, η)| |φx (x + ξ, y)| dξ 1+α +
|z|61 0 |z| 2 q p
Lx Lη Ly
Z

dz
+ C |ψ(x + z, y, η)| 1+α kφ(ξ, y)kL∞
|z|>1 |z| ξ
2 q p
Lx Lη Ly
6 C(kφ(x, y)kL∞ s + kφx (x, y)kL∞ Ls )×
x Ly
Z x y

dz
× kψ(x + z, y, η)kLqη Lry α
R |z| (1 + |z|) 2
Lx

6 C kφkL∞ s + kφx kL∞ Ls kψk 2 q r .
x Ly x y Lx Lη Ly
The commutators [|∂j |α ]Hj , φ are estimated in the same way. Lemma 2.1 is
proved. 2
We define a special operator S(ϕ) = X(ϕ)Y(ϕ), where

X(ϕ) = cosh(ϕ1 ) + isinh(ϕ1 )Hx , Y(ϕ) = cosh(ϕ2 ) + isinh(ϕ2 )Hy ,
the vector of the real-valued functions ϕ(t, x, y) = (ϕ1 (t, x), ϕ2 (t, y)), is such that
its components ϕj (t, xj ) ∈ L∞ (0, T ; H2,0,∞ ∞
x,j ) ∩ C ([0, T ]; Lxj ) are positive. From
1
this definition we easily see that the operator S acts continuously from L2 to L2
with the following estimate
kS(ϕ)ψk 6 2 exp(kϕk∞ )kψk,
where kϕk∞ = kϕ1 kL∞ x
+ kϕ2 kL∞
y
.
The inverse operator X (ϕ) = (1 + itanh(ϕ1 )Hx )−1 cosh(ϕ
−1 1
1)
also exists and is
continuous
kX−1 (ϕ)ψk 6 (1 − tanh(kϕk∞ ))−1 kψk 6 exp(kϕk∞ )kψk. (2.2)
The same is true for the operator Y.
The operator S helps us to obtain a smoothing property of the Schrödinger-
type equation (2.1) by virtue of the usual energy estimates. In the next lemma
we prepare an energy estimate, involving the operator S, in which we have an
additional positive term giving us the norm of the half derivative of the unknown
function u. We assume that ϕj (xj ) is written by ωj as ϕj (xj ) = ∂j−1 (ωj2 ), so that

p
ωj (xj ) = (∂j ϕj ), and ω = (ω1 , ω2 ).
LEMMA 2.2. The following inequality
d p 2 p 2

kSuk2 + ω1 S |∂x |u + ω2 S |∂y |u
dt
6 2|Im(Su, Sf )| + Ckuk2 e2kϕk∞ kωk4∞ + kωk1,0,∞kωk∞ + kϕt k∞
is valid for the solution u of the Cauchy problem (2.1).
Proof. Multiplying Equation (2.1) by the operator S(ϕ) we get
(i∂t + 1)S(ϕ)u − [1, S(ϕ)]u − i[∂t , S(ϕ)]u = S(ϕ)f. (2.3)
Via the property (iH)2 = 1 and [∂j , S(ϕ)] = i(∂j ϕj )S(ϕ)Hj , applying the Leib-
nitz rule we obtain
X
[1, S(ϕ)] = −2i(∂j ϕj )S(ϕ)|∂j | + (∂j ϕj )2 S(ϕ) + i(∂j2 ϕj )S(ϕ)Hj .
j =1,2
P
Similarly we have [∂t , S(ϕ)] = j =1,2 i(∂t ϕj )S(ϕ)Hj . Therefore (2.3) yields
(i∂t + 1)Su + Mu = Ru + Sf, (2.4)
where
X X
M = 2i (∂j ϕj )S(ϕ)|∂j | = 2i ωj2 S(ϕ)|∂j |
j =1,2 j =1,2
since |∂j | = −∂j Hj and the remainder term

X
R= (ωj4 S(ϕ) + 2iωj (∂j ωj )S(ϕ)Hj − (∂t ϕj )S(ϕ)Hj )
j =1,2
is a bounded operator since

kRuk 6 4kuk exp(kϕk∞ ) kωk4∞ + kωk1,0,∞kωk∞ + kϕt k∞ . (2.5)
Now we apply the usual energy method to (2.4) (i.e., we multiply (2.4) by S(ϕ)u
integrate over R2 and take the imaginary part of the result) to get
1d
kSuk2 + Im(Su, Mu) 6 |(Su, Ru)| + |Im(Su, Sf )|. (2.6)
2 dt
Then using the estimates of Lemma 2.1 we obtain
Im(Su, Mu)
X X
=2 Su, ωj2 S|∂j |u = 2 ωj Su, ∂j ωj SHj u − [∂j , ωj S]Hj u
j =1,2 j =1,2
X p p
= −2 ωj S |∂j |u + |∂j |, ωj S u,
j =1,2
p p X
− ωj S |∂j |u + |∂j |Hj , ωj S Hj u − 2 (ωj Su, [∂j , ωj S]Hj u)
j =1,2
X p p p
>2 ωj S |∂j |u 2 − ωj S |∂j u |∂j |, ωj S u +
j =1,2
p p p
+ |∂j |Hj , ωj S Hj u − |∂j |, ωj S u |∂j |Hj , ωj S Hj u −

− 2|(ωj Su, [∂j , ωj S]Hj u)|
X p
> ωj S |∂j |u 2 − Ckuk2 exp(2kϕk∞ ) kωk4 + kωk∞ kωk1,0,∞ .
∞
j =1,2
(2.7)
We have the result of the lemma from (2.5)–(2.7). 2
In the next lemma we prepare some estimates of different terms appearing in

the nonlinearity.
LEMMA 2.3. We have the following estimates
|(Su, Sφψ∂x u)|
p 2 p 2
6 2 exp(2kϕk∞ ) kφkL∞ S |∂x |u + kψkL∞ S |∂x |u +

y y
+ Ckuk2 exp(6kϕk∞ ) kφk21,0,∞ + kψk21,0,∞ 1 + kϕk21,0,∞ ,
|(Su, S∂y−1 (φψ)∂x u)|

p 2 p 2
6 4 exp(4kϕk∞ ) kφkL2y S |∂x |u + kψkL2y S |∂x |u +
+ Ckuk2 exp(6kϕk∞ ) kφk2L∞ L2 + kφx k2L∞ L2 +

x y x y
+ kψkL∞ L2 + kψx kL∞ L2 1 + kϕk1,0,∞

2 2 2
x y x y
and
|(Su, Sφ∂y−1 (ψ∂x w)|
p 2 p 2
6 4 exp(4kϕk∞ ) kφkL2y S |∂x |u + kψkL2y S |∂x |w +

+ C kuk2 + kwk2 exp(6kϕk∞ ) kφk2L∞ L2 + kφx k2L∞ L2 +

x y x y
+ kψkL∞ L2 + kψx kL∞ L2 1 + kϕk1,0,∞

2 2 2
x y x y
provided that the right-hand sides are bounded.

Proof. Since ψ∂x = ∂x ψ − ψx , we get for all 1 6 p, q 6 ∞
kXφ(x, y)ψ(x, ξ )∂x u(x, η) − Xφ(x, y)∂x ψ(x, ξ )u(x, η)kL2x L2η Lqξ Lpy
6 Ce2kϕk∞ kφkL∞ p kψx k ∞ p kuk.
x Ly Lx Ly (2.8)
Using the identity

p p
Hx φ |∂x | − |∂x |(φHx )
p p p
= − |∂x |, φ Hx + |∂x |Hx , φ − Hx |∂x |, φ ,
we write the representation
p p p p
Xφ |∂x | = |∂x |φX − |∂x |, cosh(ϕ1 )φ − i [ |∂x |, sinh(ϕ1 ) φHx +
p p
+ isinh(ϕ1 ) Hx φ |∂x | − |∂x |(φHx )
p p
= |∂x |φX − isinh(ϕ1 )Hx |∂x |, φ + R, (2.9)
where the remainder operator
p p
R(x, y) = isinh(ϕ1 ) |∂x |Hx , φ(x, y) − isinh(ϕ1 ) |∂x |, φ(x, y) Hx −
p p
− |∂x |, cosh(ϕ1 )φ(x, y) − i |∂x |, sinh(ϕ1 ) φ(x, y)Hx .
Via the estimates of Lemma 2.1 we have the estimate for R
kR(x, y)w(x, y, η)kL2x Lqη Lpy

6 C exp(kϕk∞ ) kφkL∞ r + kφx kL∞ Lr ×
x Ly

x y
× (1 + kϕk1,0,∞) kwkL2x Lqη Lsy + kHx wkL2x Lqη Lsy , (2.10)
where 1r + 1s = p1 .
Next we note that
p p p
X |∂x |Hx ψ = ψX |∂x |Hx + cosh(ϕ1 ) |∂x |Hx , ψ +
p
+ isinh(ϕ1 ) |∂x |, ψ ,
whence by Lemma 2.1 we find
p
X |∂x |Hx ψ(x, y)w(x, y, η) 2 q p
Lx Lη Ly
p
6 ψ(x, y)X |∂x |Hx w(x, y, η) L2 Lq Lp +

x η y
+ Ce2kϕk∞
kψkL∞ r + kψx kL∞ Lr kwk 2 q s ,
x Ly x y Lx Lη Ly (2.11)
where + 1s = p1 . By virtue of Lemma 2.1 we have the estimate

1
r
p p p
gX |∂x |Hx u 6 |∂x |Hx gXu + |∂x |Hx , gX u
p p
6 gY−1 S |∂x |u + |∂x |, gX u +
p
+ |∂x |Hx , gX u
p
6 exp(2kϕk∞ ) kgkL∞ S |∂x |u +
y
+ Ckuk exp(2kϕk∞ )kgk1,0,∞ (1 + kϕk1,0,∞ ). (2.12)

Denoting the conjugate operator Y∗ = cosh(ϕ2 ) − iHy sinh(ϕ2 ) in L2y we obtain by

virtue of (2.8) with p = q = ∞
|(Su, Sφψ∂x u)| = |(Y∗ Su, Xφψ∂x u)|
6 |(Y∗ Su, Xφ∂x ψu)| + kY∗ SukkXφψx uk
6 |(Y∗ Su, Xφ∂x ψu)| + Ce2kϕk∞ kφk∞ kψk1,0,∞ kuk2 .
We consider the first term of the right-hand side of the √

above. Via the estimate
(2.10) with q = r = ∞, p = s = 2 and w(x, y, η) = |∂x |Hx ψ(x, y)u(x, y),
and (2.9) we get
∗
(Y Su, Xφ∂x ψu)
p p
= Y∗ Su, Xφ |∂x | |∂x |Hx ψu
p p
6 φ |∂x |Y∗ Su, X |∂x |Hx ψu +
p p
+ Hx sinh(ϕ1 )Y∗ Su, |∂x |, φ |∂x |Hx ψu +
p
+ kY∗ Suk R |∂x |Hx ψu
p p
6 φ |∂x |Y∗ Su + Ckuke3kϕk∞ kφk1,0,∞ (1 + kϕk1,0,∞ ) X |∂x |Hx ψu ,
whence using (2.11) with w(x, y, η) = u(x, y), and (2.12) with g = φ, q = r =
∞, p = s = 2, we find
∗
Y Su, Xφ∂x ψu
p
6 φY∗ S |∂x |u + Ckuke3kϕk∞ kφk1,0,∞(1 + kϕk1,0,∞ ) ×
p
× ekϕk∞ kψkL∞ y
S |∂x |u + Ckuke3kϕk∞ kψk1,0,∞ (1 + kϕk1,0,∞)
p 2 p
6 2e4kϕk∞ kφkL∞ S |∂ x |u + kψkL∞ S |∂x |u 2

y y
+ Ckuk2 e6kϕk∞ kφk21,0,∞ + kψk21,0,∞ 1 + kϕk21,0,∞ .

Thus the first estimate of the lemma is proved. In view of (2.8) with p = q = 2 we
have

Su, S∂ −1 (φψ)∂x u = Y∗ Su, X∂ −1 (φψ)∂x u
y y

6 Y∗ Su, X∂y−1 (φ∂x ψ)u + kY∗ Suk X∂y−1 (φψx )u

6 Y∗ Su, X∂y−1 (φ∂x ψ)u + Ce2kϕk∞ kφkL∞ 2 ×
x Ly

× kψkL∞ 2 + kψx kL∞ L2 kuk .
x Ly x y
2
Applying inequality (2.10) with p = 1, q = r = s = 2 and w(x, y, η) =

√
|∂x |Hx ψ(x, y)u(x, η) we get
∗
Y Su, X∂ −1 (φ∂x ψ)u
y
p p
6 |∂x |Y∗ Su, ∂y−1 φX |∂x |Hx ψ u +
p p
+ Hx sinh(ϕ1 )Y∗ Su, ∂y−1 |∂x |, φ |∂x |Hx ψ u +
p
+ kY∗ Suk R(x, y) |∂x |Hx ψ(x, y)u(x, η) L2 L2 L1
p p
x η y

6 φ(x, y) |∂x |Y∗ Su(x, η) L2 L2 L2 X |∂x |Hx ψ(x, y)u(x, η) L2 L2 L2 +
x η y x η y
+ Ckuke 4kϕk∞
(1 + kϕk1,0,∞ )×
p p
× |∂x |, φ(x, y) |∂x |Hx ψ(x, y)u(x, η) L2 L2 L1 +
p
x η y

+e 2kϕk∞
kφkL∞ L2 + kφx kL∞ L2 X |∂ x |Hx ψ(x, y)u(x, η) 2 2 2
x y x y L L L
x η y
whence using estimate (2.11) with p = q = r = 2, s = ∞ and w(x, y, η) =

u(x, η), and after that (2.12) with g = kφkL2y , we find
∗
Y Su, X∂ −1 (φ∂x ψ)u
y
p
6 e2kϕk∞ kφkL2y S |∂x |u +

+ Ckuke3kϕk∞ kφkL∞ L2 + kφx kL∞ L2 (1 + kϕk1,0,∞ ) ×
p
x y x y
2kϕk∞

× e kψkL2y S |∂x |u +

+ Ckuke 3kϕk∞
kψkL∞ L
x y
2 + kψx kL∞ L2 (1 + kϕk1,0,∞ )
x y
4kϕk∞
p 2 p 2
6 4e kφkL2y S |∂x |u + kψkL2y S |∂x |u + Ckuk2 e6kϕk∞ ×

× (kφkL∞ 2 + kφx kL∞ L2 + kψkL∞ L2 + kψx kL∞ L2 )(1 + kϕk1,0,∞ ). (2.13)
x Ly x y x y x y
2
This shows the second estimate in the lemma. Via estimate (2.8) with p = q = 2
we have

Su, Sφ∂ −1 (ψ∂x w)
y

= Y∗ Su, Xφ∂y−1 (ψ∂x w)

6 Y∗ Su, Xφ∂y−1 (∂x ψw) + kY∗ Suk Xφ∂y−1 (ψx w)

6 Y∗ Su, Xφ∂y−1 (∂x ψw) + Ce2kϕk∞ kφkL∞ 2×
x Ly

× kψkL∞ 2 + kψx kL∞ L2 kukkwk.
x Ly x y
√ view of estimate (2.10) with p = r = 2, q = 1, s = ∞ and w(x, y, η) =

In
|∂x |Hx ψ(x, y)w(x, η) we obtain
∗
Y Su, Xφ∂ −1 (∂x ψw)
y
p p
6 φ |∂x |Y∗ Su, ∂y−1 X |∂x |Hx ψw +
p p
+ Hx sinh(ϕ1 )Y∗ Su, |∂x |, φ ∂y−1 |∂x |Hx ψw +
p
+ kY∗ Suk R(x, y) |∂x |Hx ψ(x, η)w(x, η) L2 L1 L2
p p
x η y

6 φ(x, y) |∂x |Y Su(x, y) L2 L1 X |∂x |Hx ψ(x, y)w(x, y) L2 L1 +
∗
x y x y
+ Ckuke4kϕk∞ (1 + kϕk1,0,∞ )×
p p
× |∂x |, φ(x, y) |∂x |Hx ψ(x, η)w(x, η) L2 L1 L2 +
p
x η y

+e 2kϕk∞
kφkL2y L∞ + kφx kL2y L∞ X |∂x |Hx ψ(x, y)w(x, y) L2 L1 .
x x x y
In the same way as in the proof of (2.13), by virtue of estimate (2.11) with p =
1, q = ∞, r = s = 2 and w(x, y, η) = w(x, y), and after that using inequality
(2.12) with g = kφkL2y and g = kψkL2y , we get
∗
Y Su, Xφ∂ −1 (∂x ψw)
y
p
6 e2kϕk∞ kφkL2y S |∂x |u +

+ Ckuke3kϕk∞ kφkL∞ x Ly
2 + kφx kL∞ L2 (1 + kϕk1,0,∞ ) ×
p
x y

× e2kϕk∞ kψkL2y S |∂x |w +

+ Ckwke3kϕk∞ kψkL∞ x Ly
2 + kψx kL∞ L2 (1 + kϕk1,0,∞ )
x y
p 2 p 2
6 4e4kϕk∞ kφkL2y S |∂x |u + kψkL2y S |∂x |w +

+ Ce6kϕk∞ kuk2 + kwk2 kφkL∞ 2 + kφx kL∞ L2 + kψkL∞ L2 +
x Ly

x y x y
+ kψx kL∞
x Ly
2 1 + kϕk1,0,∞ .
2
(2.14)
Thus the third estimate of the lemma follows from (2.14). Lemma 2.3 is proved. 2
3. Local Existence
In this section we prove Theorems 1.1 and 1.2, which state the local-in-time exis-
tence of solutions to the Cauchy problem (1.1). The main point is the use of the
smoothing operator S and the energy type estimate of Lemma 2.2.
Proof of Theorem 1.1. We put w = (1 − 1)u, then the vector v is defined by

the function w as follows v = ((1 − 1)−1 w, (1 − 1)−1 ∂x w, (1 − 1)−1 ∂y w, (1 −
1)−1 w̄, (1 − 1)−1 ∂x w̄, (1 − 1)−1 ∂y w̄). We get from (1.1)

Lw = G(v, w) + R(v, w), (x, y, t) ∈ R3 ,
(3.1)
w(0, x, y) = (1 − 1)u0 (x, y), (x, y) ∈ R2 ,
where L = i∂t + 1, the main part
G(v, w)
X
= (N1 (v))uj ∂j w + (N1 (v))ūj ∂j w̄
j =x,y

+ wx ∂j−1 Kj(2) (v) + wy ∂j−1 Kj(3) (v) + w̄x ∂j−1 Kj(5) (v) + w̄y ∂j−1 Kj(6) (v) +
+ v, ∂x−1 (Kx (v))uy wy + ∂x−1 (Kx (v))ūy w̄y + ∂y−1 (Ky (v))ux wx +

+ ∂y−1 (Ky (v))ūx w̄x
depends linearly on the derivatives of w and the remainder operator

X X
R(v, w) = N (v) − (N (v))v (k)v (l) ∂j v (k) ∂j v (l) + (N1 (v))u w +
j =x,y 16k,l 66
X
+ (N1 (v))ū w̄ + w∂j−1 Kj(1) (v) + w̄∂j−1 Kj(4) (v)
j =x,y
+ v, ∂x−1 (Kx (v))u w+ ∂x−1 (Kx (v))ū w̄ +

+ ∂y−1 (Ky (v))u w + ∂y−1 (Ky (v))ū w̄
can be easily estimated

kR(v, w)k2,0 6 C kwk2∞ + kwk2L2 H1,0,∞ + kwk2L2 H1,0,∞ kwk2,0
y x x y
6 Ckwk32,0 (3.2)
if the solution is sufficiently small kwk 6 ε 6 1. Now let us consider the linearized
version of Equation (3.1)

Lw = G(ṽ, w) + R(ṽ, w̃), (x, y, t) ∈ R3 ,
(3.3)
w(0, x, y) = (1 − 1)u0 (x, y), (x, y) ∈ R2 ,
where the vector ṽ = ((1 − 1)−1 w̃, (1 − 1)−1 w̃x , (1 − 1)−1 w̃y ,
¯ (1 − 1)−1 w̃
(1 − 1)−1 w̃, ¯ x , (1 − 1)−1 w̃
¯ y ) is defined by the known function w̃,
from the ball
n
B = w̃ ∈ C1 [0, T ]; L2 (R2 ) : sup kw̃k 6 2ε, sup kw̃k2,0 6 2ρ,
t ∈[0,T ] t ∈[0,T ]
−1 −1 o
sup ∂t ∂x kw̃kL2 L∞ + ∂t ∂y kw̃kL2 L∞ 6 4ρ 2 ,
2 2
y x x y
t ∈[0,T ]
where ε = ku0 k2,0 , ρ = ku0 k4,0 . Thus the Cauchy problem (3.3) defines a map-
ping M : w = M(w̃). First let us show that there exists a time T > 0, such that the
mapping M transforms the closed ball B into itself provided that the value ε > 0 is
sufficiently small. After that we prove that there exists a time T > 0, such that M
is a contraction mapping in the norm supt ∈[0,T ] k · k under the constraint that it acts
on the subspace B, and ε > 0 is sufficiently small. By the usual energy method we
have from Equation (3.3)
Z t
kw(t)k 6 ku0 k2,0 + Cρ
2 2 2
kw(τ )k22,0 dτ. (3.4)
0
In order to obtain the estimates of the norm supt ∈[0,T ] kw(t)k2,0 we consider the
function h = 1w. We use the smoothing operators S(ϕ) = X(ϕ)Y(ϕ), intro-
duced in Section 2, where X(ϕ) = cosh(ϕ1 ) + isinh(ϕ1 )Hx , Y(ϕ) = cosh(ϕ2 ) +
isinh(ϕ2 )Hy , the components of the vector ϕ(t, x, y) = (ϕ1 (t, x), ϕ2 (t, y)) are
defined as ϕ1 (t, x) = 1ε ∂x−1 kw̃(t, x, y)k2L2 and ϕ2 (t, y) = 1ε ∂y−1 kw̃(t, x, y)k2L2 ,
y x
so that ϕj (t, xj ) ∈ L∞ (0, T ; H2,0,∞

xj ) ∩ C1 ([0, T ]; L∞
xj ) and we denote as in Sec-
tion 2 ω1 (t, x) = √ε kw̃(t, x, y)kL2y and ω2 (t, y) = √1ε kw̃(t, x, y)kL2x . Therefore
1
applying Lemma 2.2 we obtain the energy type inequality

d p 2 p 2
kShk2 + ω1 S |∂x |h + ω2 S |∂y |h
dt
6 2|Im(Sh, S1G(ṽ, w))| + 2|Im(Sh, S1R(ṽ, w̃))| +

+ Ce2kϕk∞ kωk4∞ + kωk1,0,∞kωk∞ + kϕt k∞ khk2 . (3.5)
Since w̃ ∈ B we have kωk∞ 6 √2 (kw̃kL∞ L2
ε x y
+ kw̃kL∞ 2 ) 6 1, kωk1,0,∞ 6
y Lx
√2 (kw̃kL∞ L2
ε x y
+ kw̃x kL∞ 2 + kw̃kL∞ L2 + kw̃y kL∞ L2 ) 6
x Ly y x y x
6ρ
√
ε
, kϕk∞ 6 1ε kw̃k2 6 1
and kϕt k∞ 6 1ε (k∂t ∂x−1 kw̃k2L2 kL∞
x
+ k∂t ∂y−1 kw̃k2L2 kL∞
y
) 6 Cρ 2 /ε. In view of (3.2)
y x
we get
|Im(Sh, S1R(ṽ, w̃))| 6 e2kϕk∞ khk kR(ṽ, w̃)k2,0 6 Cρ 3 khk. (3.6)
From the definition of the main part G(ṽ, φ) we obtain k1G(ṽ, w) − G(ṽ, 1w)k 6
Cρ 2 kwk2,0 . Thus we have
|Im(Sh, S1G(ṽ, ω̃))| 6 |Im(Sh, SG(ṽ, h))| + Cρ 2 kwk2,0 khk. (3.7)
To estimate the main term Im(Sh, SG(ṽ, h)) in the right-hand side of (3.7) we
apply Lemma 2.3. Since N1 (ṽ) is cubic and (N1 (ṽ))uj is quadratic with respect to
|(N (ṽ)) |
the vector ṽ in the neighborhood of the origin we have k 1|ṽ| ux kL∞ 6 CkṽkL∞ 6
√
y y
|(N1 (ṽ))ux |
εCω1 and k |ṽ| k1,0,∞ 6 Ckṽk1,0,∞ 6 Cρ, therefore applying the first esti-
and ϕ = ( 1ε ∂x−1 kw̃(t, x, y) k2L2 ,
(N1 )ux
mate of Lemma 2.3 with h = u, φ = |ṽ|, ψ = |ṽ| y
1 −1
∂ kw̃(t, x, y)k2L2 )
ε y
we get
x

Sh, S(N1 )u ∂x h = Sh, S|ṽ| (N1 )ux ∂x h
x
|ṽ|
p 2
6 Cε ω1 S |∂x |h + Cρ 2 (1 + ρ)khk2 .
The value |(Sh, S(N1 )uy ∂y h)| is estimated in the same way. Using the fact that
Kj(l) (ṽ) are quadratic and hence |Kj(l) (ṽ)|/|ṽ| are linear with respect to the vector
ṽ in the neighborhood of the origin, we have
(l)
|Kj (ṽ)| √

|ṽ| 2 6 CkṽkL2y 6 εCω1
Ly
and
(l) (l)
|Kj (ṽ)| |Kj (ṽ)|

|ṽ| ∞ 2 + |ṽ|
Lx Ly x L∞ 2
x Ly
6 CkṽkL∞ 2 + Ckṽx kL∞ L2 6 Cρ;

x Ly x y
therefore, by the second estimate of Lemma 2.3 we obtain

p
Sh, Shx ∂ −1 K (2) (ṽ) 6 Cε ω1 S |∂x |h 2 + Cρ 2 (1 + ρ)khk2 .
y y
The estimate of the summand |(Sh, Shy ∂x−1 Kx(3) )| is obtained in the same way
as in the above. Finally, since the functions (Ky(l) (ṽ))ux are linear with respect
√
to the vector ṽ, we have the estimates k(Ky(l) (ṽ))ux kL2y 6 CkṽkL2y 6 εCω1
and k(Ky(l) (ṽ))ux kL∞ 2 + k((Ky (ṽ))ux )x kL∞ L2 6 CkṽkL∞ L2 + kṽx kL∞ L2 6 Cρ,
x Ly
(l)
x y x y x y
whence by the third estimate of Lemma 2.3 we have
p
Sh, Sv (l) ∂ −1 K (l) (ṽ) hx 6 Cε ω1 S |∂x |h 2 + Cρ 2 (1 + ρ)khk2 ,
y y ux
for l = 1, . . . , 6. The terms |(Sh, Sv (l)∂x−1 (Kx(l) )uy hy )| are estimated analogously.
Thus we get for the second term G:
p 2 p 2
|Im(Sh, SG(ṽ, h))| 6 Cε ω1 S |∂x |h + Cε ω2 S |∂x |h +
+ Cρ 2 (1 + ρ)khk2 . (3.8)
Substitution of (3.6)–(3.8) into (3.5) yields

d p 2 p 2
kShk2 + (1 − Cε) ω1 S |∂x |h + (1 − Cε) ω2 S |∂x |h
dt
6 Cρ 2 kwk + Cρ 2 khk + C 1 + ρ 3 khk2 6 C 1 + ρ 3 + kwk22,0 . (3.9)
Integration of (3.4) and (3.9) gives us desired estimates kwk 6 2ε and kwk2,0 6
kwk + e2kϕk∞ kShk 6 2ρ, if the time interval T > 0 is sufficiently small.
Now directly from system (3.3) we see that
Z
−1
∂t ∂ kw̃k2 2 ∞ = 2 ∂ −1
Re(w̄w t ) dy

x Ly Lx x ∞
Lx
6 kw̄wx kL∞ 1 + 2kwkkG(ṽ, w)k + 2kwkkR(ṽ w̃)k
x Ly
6 kwkL∞ 2 kwx kL∞ L2 + Cερkwk2,0 + Cεkw̃k2,0 6 4ρ ,

x Ly x y
2 2
if ε > 0 is sufficiently small. Thus the mapping M transforms the ball B into itself.
Let us now show that M is a contraction mapping in the norm supt ∈[0,T ] k · k. Let
ω† satisfy Equation (3.3) with the known function w̃ † ∈ B instead of w̃. Then for
the difference g = w † − w we get

 Lg = G(ṽ † , g) + G(ṽ † , w) − G(ṽ, w) +
+ R(ṽ † , w̃ † ) − R(ṽ, w̃), (x, y, t) ∈ R3 , (3.10)

g(0, x, y) = 0, (x, y) ∈ R .
2
Denoting g̃ = w † − w̃ we easily get kG(ṽ † , w) − G(ṽ, w)k 6 Cρ 2 kg̃k and

kR(ṽ † , w̃ † ) − R(ṽ, w̃)k 6 Cρ 2 kg̃k. From (3.10) we find the estimate for g in
the same way as in the proof of (3.9):

d p 2 p 2
kSgk2 + (1 − Cε) ω1 S |∂x |g + (1 − Cε) ω2 S |∂x |g
dt
6 C 1 + ρ 2 kg̃k, (3.11)
whence integrating with respect to time t, we get the desired estimate

1
sup kgk 6 CT (1 + ρ 2 ) sup kg̃k 6 sup kg̃k
t ∈[0,T ] t ∈[0,T ] 2 t ∈[0,T ]
if we choose the time interval T > 0 to be sufficiently small. Thus the trans-
formation M is a contraction mapping. Therefore there exists a unique solution
u ∈ C([0, T ]; H3,0 ) ∩ L∞ (0, T ; H4,0 ) of the Cauchy problem (1.1). Theorem 1.1
is proved. 2
Proof of Theorem 1.2. Now we assume that the local nonlinear part N1 =
N1 (u, ū), Kj(l) = Kj(l) (u, ū), j = x, y, l = 2, 3, 5, 6, the functions Kj(1) and
Kj(4) depend linearly on u, ū, so that the derivatives (Kx )uy , (Kx )ūy , (Ky )ux
and (Ky )ūx depend on u, ū only. Considering a linearized version of Equation
(1.1) similar to Equation (3.3), we take the known function ũ to be fixed from
the ball ũ ∈ B = {ũ ∈ C1 ([0, T ]; L2 ) : supt ∈[0,T ] kũk 6 2ε, supt ∈[0,T ] kũk2,0 6
2ρ, supt ∈[0,T ] (k∂t ∂x−1 kũk2L2 kL∞
x
+k∂t ∂y−1 kũk2L2 kL∞
y
) 6 4ρ 2 }, where ε = ku0 k, ρ =
y x
ku0 k2,0 . Now the local part N1 does not give us any derivative loss, so we can
choose the vector ϕ(t, x, y) = (ϕ1 (t, x), ϕ2 (t, y)) in the definition of the smooth-
ing operator S to be more simple ϕ1 (t, x) = 1ε ∂x−1 kũ(t, x, y)k2L2 and ϕ2 (t, y) =
y
1 −1
∂ kũ(t, x, y)k2L2 .
ε y
All the other costructions are similar to the proof of Theo-
x
rem 1.1, so we leave them to the reader. 2
4. Global Existence
Repeating the proof of Theorem 1.1 and applying the operators Jx , Jy we can
easily obtain an additional estimate of U(−t)u(t) in the weighted Sobolev space
H4,1 . (See also the proof of Theorem 1.3 below.)
THEOREM 4.1. Let the initial data u0 ∈ H4,1 and the norm ε = ku0 k2,0 of
the initial data is sufficiently small, then there exists a time T > 0 and a unique
solution U(−t)u ∈ C([0, T ]; H3,0 ) ∩ L∞ (0, T ; H4,1 ) to the Cauchy problem (1.1).
Moreover, we can choose the existence time T > 0 so that the following es-
timate supt ∈[0,T ] ku(t)k4,0 6 2ku0 k4,0 is valid for the solution u of the Cauchy
problem (1.1).
Proof of Theorem 1.3. Let us prove the following a-priori estimates for the
function w = (1 − 1)u
(1 + t)−γ
√ kwk2,0 < 2ε, (4.1)
kwk + 1 + t kwkL∞ L
x y
2 + kwx kL∞ L2 + kwkL∞ L2 + kwy kL∞ L2 < 100ε
x y y x y x
for all t > 0, where γ > 0 is small. We show (4.1) by arguments of the contra-
diction. We assume that there exists a time T > 0 such that estimates (4.1) are
violated in the interval [0, T ]. By Theorem 4.1 we have the following estimates
(1 + t)−γ
√ kwk2,0 6 4ε, (4.2)
kwk + 1 + t kwkL∞ 2 + kwx kL∞ L2 + kwkL∞ L2 + kwy kL∞ L2 6 100ε
x Ly x y y x y x
for all t ∈ [0, T ]. To estimate the norms kwk2,0 and kU(−t)wk2,1 we multiply
Equation (3.1) by the operators A = 1−1, and A = (1−1)Jl , where l = x, y and
then we use the same smoothing operator S(ϕ) as in Section 3. Applying Lemma
2.2 we obtain the estimates for the functions h = Aw
d p 2 p 2
kShk2 + ω1 S |∂x |h + ω2 S |∂y |h
dt
6 2|Im(Sh, SAG(v, w))| + 2|Im(Sh, SAR(v, w))|+

+ Ce2kϕk∞ kωk4∞ + kωk1,0,∞kωk∞ + kϕt k∞ khk2 . (4.3)
√
Now by virtue of (4.2) we get the estimates kωk1,0,∞ 6 C ε(1 + t)−1/2 , kϕk∞ 6
1 and kϕt k∞ 6 Cε(1 + t)−1 . Since kφx k2L4 L2 6 kφkL∞ 2 kφk2,0 in view of (4.2)
x Ly
x y
we also have the estimate
|Im(Sh, SAR(v, ω))| 6 Cε 3 (1 + t)−1+2γ + Cε(1 + t)−1 khk2 . (4.4)
Using the identities φ̄ψx = 2it 1
(φ̄Jx ψ − (Jx φ)ψ) − φ̄x ψ and φψx = 2it 1
(φJx ψ −
(Jx φ)ψ) + φx ψ with φ = v (l) , ψ = wx we can estimate the terms of the
form, for example, v̄ (l) wxx Jx v (k) in the operator G(v, w), so we obtain the estimate
kAG(v, w) − G(v, Aw)k 6 Cε 2 (1 + t)−1+γ + Cε(1 + t)−1 khk. Thus we get
|Im(Sh, SAG(v, w))|
6 |Im(Sh, SG(v, h))| + Cε 3 (1 + t)−1+2γ + Cε(1 + t)−1 khk2 . (4.5)
Then, as in Section 3, we obtain via (4.2)
p 2 p 2
|Im(Sh, SG(v, h))| 6 Cε ω1 S |∂x kh + Cε ω2 S |∂y kh +
+ Cε 3 (1 + t)−1+2γ + Cε(1 + t)−1 khk2 . (4.6)
Substitution of (4.4)–(4.6) into (4.3) yields
d p 2 p 2
kShk2 + (1 − Cε) ω1 S |∂x kh + (1 − Cε) ω2 S |∂y kh
dt
6 Cε(1 + t)−1 kShk2 + Cε 3 (1 + t)−1+2γ . (4.7)
Integration of (4.7) gives us the desired estimate kwk2,0 < 2ε(1 + t)γ and also
kU(−t)wk2,1 < Cε(1 + t)γ for all t ∈ [0, T ].
In order to estimate the norms kwk, kwkL∞ 2 , kwx kL∞ L2 , kwkL∞ L2 and
x Ly x y y x
1 iτ ξ 2 +iτ η2
kwy kL∞ 2 we change a dependent variable u(t, x, y) =
y Lx τ
e U (τ, ξ, η),
1 iτ ξ 2 +iτ η2
where ξ = 2τx , η = 2τy , τ = 1 + t, whence ux (t, x, y) = τ e Iξ U (τ, ξ, η),
and uy (t, x, y) = τ1 eiτ ξ +iτ η Iη U (τ, ξ, η)
2 2
with operators Iξ = iξ + 2τ1 ∂ξ , and
Iη = iη + 2τ1 ∂η . Taking into account the complex-conjugate structure of the
nonlinearity in Equation (1.1) we obtain
 V V
 1 −1

 iU + 1U + τ N + τ V , ∂ K +
 τ 2 1 ξ ξ
4τ V τ τ
(4.8)

 + τ V , ∂η−1 Kη = 0, (ξ, η) ∈ R2 , τ > 1,

 τ
U (1, ξ, η) = u0 (ξ, η)e−iξ −iη ,
2 2
(ξ, η) ∈ R2 ,
where the vectors V = (U, Iξ U, Iη U, U , Iξ U , Iη U ). Note also that
i iτ ξ 2 +iτ η2
(Jx + 2i∂x )u(t, x, y) = (x + 2iτ ∂x )u(t, x, y) = e Uξ (τ, ξ, η)
τ
and similarly we have (Jy + 2i∂y )u(t, x, y) = τi eiτ ξ +iτ η Uη (τ, ξ, η).
2 2
We denote W (t, ξ, η) = (1 − Iξ2 − Iη2 )U (t, ξ, η) = τ eiτ ξ −iτ η w, then we

2 2
have the following relations kwk = kW k, kwkL∞ x Ly

2 = √ kW kL∞ L2 and similarly
1
τ ξ η
kwx kL∞ 2 = √ kIξ W kL∞ L2 . Let us obtain the main term in the nonlinearity of
x Ly
1
τ ξ η
Equation (4.3) which decays slowly with time. Using the identity
1 1
ϕ̄Iξ ψ = (ϕ̄Iξ ψ − ϕIξ ϕ) + (ϕ̄ψ)ξ , (4.9)
2 4τ
where we take ϕ = ψ = U , we write the following representation iλj(2) U Iξ U =
i (2)
λ ((U Iξ U −U Iξ U )+ 2τ1 (|U |2 )ξ ) = −λ(2)
2 j
i (2)
j Im(U Iξ U )+ 4τ λj (|U | )ξ . The term
2
iλj U Iξ U has a similar representation. In the same manner we get iλj(4) U Iη U =

(3)
−λj(4)Im(U Iη U ) + 4τi λ(4) (5)

j (|U | )η , and analogously iλj U Iη U is considered. Ap-
2
plying the identity

1
ϕ̄Iξ ψ = −ψIξ ϕ + (ϕ̄ψ)ξ (4.10)
2τ
with ϕ = Iη U , ψ = U , we get Iξ U Iη U = −U Iξ Iη U + 2τ1 (U Iη U )ξ whence we
find λ(8) (8) 1 (8)
j Iξ U Iη U = λj Re(Iξ U Iη U ) + 2τ λj ((U Iη U )ξ − (U Iξ U )η ). Therefore
we have the following representation

Kj(1) = λj(1) |U |2 − λj(2) − λj(3) Im(U Iξ U ) − λj(4) − λj(5) Im(U Iη U ) +
(9)
+ λj(6) |Iξ U |2 + λj(7)|Iη U |2 + λ(8)
j + λj Re(Iξ U Iη U ) +
i (2) i (4)
+ λj − λj(3) |U |2 ξ + λj − λ(5)j |U |2 η +
4τ 4τ
1 (8) V
+ λj − λj(9) (U Iη U )ξ − (U Iξ U )η + Bj(1) . (4.11)
2τ τ
Using the identity ψIη φ = iηφψ + 2τ1 ψφη we get Iη U ∂ξ−1 Kξ(3) = U ∂ξ−1 iηKξ(3) +
1
U ∂ −1 Kξ(3). By virtue of (4.9), (4.10) as above we will extract the real part in
2τ η ξ
each term of iηKξ(3) . We have −µ(1) η|U |2 = −µ(1) Im(U Iη U ) − 2τi µ(1) (|U |2 )η ,
then we get iµ(2) ηU Iξ U = µ(2) Re(U Iη Iξ U )− 4τ1 µ(2) (U Iξ U )η − 4τ1 µ(2) (U Iη U )ξ .
The term iµ(3) ηU Iξ U has a similar representation. Then we obtain iµ(4) ηU Iη U =
µ(4) |Iη U |2 − 2τ1 µ(4) Uη Iη U , and the summand iµ(5) ηU Iη U is considered analo-
gously. Furthermore we get −µ(6) η|Iξ U |2 = −µ(6) Im(Iξ U Iη Iξ U )− 2τ1 µ(6) |Iξ U |η ,
and −µ(8) ηIη U Iξ U = −µ(8) Im(Iη U Iη Iξ U )+ 4τi µ(8) (|Iη U |)ξ − 2τi µ(8) Iη U Iξ Uη .
The values −µ(7) η|Iη U |2 and −µ(9) ηIξ U Iη U are considered in the same manner.
Therefore we obtain

(Iη U )∂ξ−1 Kξ(3) = U ∂ξ−1 −µ(1) Im(U Iη U ) + µ(2) + µ(3) Re(U Iη Iξ U )+

+ µ(4) + µ(5) |Iη U |2 − µ(6) Im(Iξ U Iη Iξ U )−

− µ(7) Im Iη U Iη2 U − µ(8) − µ(9) Im(Iη U Iη Iξ U )−
1
− 2iµ(1) (|U |2 )η − µ(2) U Iξ U η − µ(2) (U Iη U )ξ − µ(3) (U Iξ U )η −
4τ
− µ(3) (U Iη U )ξ − 2µ(4) Uη Iη U − 2µ(5) U η Iη U − 2µ(6) |Iξ U |2 η −

− 2µ(7) |Iη U |2 η + i µ(8) − µ(9) (|Iη U |)ξ − 2iµ(8) Iη U Iξ Uη +
1
+ 2iµ(9) Iξ U Iη Uη + U ∂ξ−1 iηBξ(3) + Uη ∂ξ−1 Kξ(3) . (4.12)
2τ
The summand (Iξ U )∂η−1 Kη(2) can be represented in the same manner. Thus via
(4.11) and (4.12) we get for the nonlinearity τ N ( Vτ ) = τ1 QU + P (U ) + τ N1 ( Vτ ),
where the coefficient at the main term is the real valued function
X (3) (5)
Q = ∂j−1 λ(1) (2) (4)
j |U | − λj − λj Im(U Iξ U ) − λj − λj Im(U Iη U ) +
2
j =x,y
(9)
+ λ(6) (7) (8)
j |Iξ U | + λj |Iη U | + λj + λj Re(Iξ U Iη U ) +
2 2

+ ∂ξ−1 − µ(1) Im(U Iη U ) + µ(2) + µ(3) Re(U Iη Iξ U ) +

+ µ(4) + µ(5) |Iη U |2 − µ(6) Im(Iξ U Iη Iξ U ) − µ(7) Im(Iη U Iη2 U ) −

− µ(8) − µ(9) Im(Iη U Iη Iξ U ) + ∂η−1 − ν (1) Im(U Iξ U ) +

+ ν (2) + ν (3) |Iξ U |2 + ν (4) + ν (5) Re(U Iη Iξ U ) − ν (6) Im Iξ U Iξ2 U −

− ν (7) Im(Iη U Iη Iξ U ) − ν (8) − ν (9) Im(Iξ U Iη Iξ U )
and the remainder operator is

ψ X −1
P (ψ) = 2
∂j i λj(2) − λ(3) j |U |2 ξ + i λj(4) − λj(5) |U |2 η +
4τ j =x,y

(8) (9) 2 (1) V
+ 2 λj − λj (U Iη U )ξ − (U Iξ U )η + τ Bj −
τ
ψ
− 2 ∂ξ−1 2iµ(1) |U |2 η − µ(2) (U Iξ U )η − µ(2) (U Iη U )ξ −
4τ
− µ(3) (U Iξ U )η − µ(3) (U Iη U )ξ − 2µ(4) Uη Iη U − 2µ(5) U η Iη U −

− 2µ(6) |Iξ U |2 η − 2µ(7) |Iη U |2 η + i µ(8) − µ(9) |Iη U | ξ −

− 2iµ(8) Iη U Iξ Uη + 2iµ(9) Iξ U Iη Uη + iηBξ(3) −
ψ
− 2 ∂η−1 2iν (1) |U |2 ξ − ν (2) (U Iξ U )ξ − ν (2) (U Iξ U )ξ −
4τ
− ν (3) (U Iξ U )ξ − ν (3) (U Iξ U )ξ − 2ν (4) Uξ Iη U − 2ν (5) U ξ Iη U −

− 2ν (6) |Iξ U |2 ξ − 2ν (7) |Iη U |2 ξ + i ν (8) − ν (9) (|Iξ U |)η −
1
− 2iν (8) Iξ U Iη Uξ + 2iν (9) Iη U Iξ Uξ + iξ Bη(2) + ψη ∂ξ−1 Kξ(3) +
2τ
1
+ ψξ ∂η−1 Kη(2) .
2τ
Now we apply the following identities Ij (φψ) = ψIj φ+ 2τ1 φ∂j ψ, and Ij (ϕψ φ̄) =
ψ φ̄Ij ϕ + ϕ φ̄Ij ψ + ϕψIj φ = ψ φ̄Ij ϕ + 2τ1 ϕ∂j (ψ φ̄), j = ξ, η to represent the
nonlinearity in the form (1 − Iξ2 − Iη2 )τ N ( Vτ ) = τ1 QW + R0 , and analogously we
can write (1 − Iξ2 − Iη2 )τ Iξ N ( Vτ ) = τ1 QIξ W + R1 , (1 − Iξ2 − Iη2 )τ Iη N ( Vτ ) =
1
τ
QIη W + R2 , where via (4.11) and (4.12) the remainder terms are

V
R0 = P (W ) + (1 − Iξ − Iη )τ N1
2 2
−
τ
1
− 2 2(Iξ U )Kξ(3) + 2(Iξ U )∂ξ ∂η−1 Kη(2) +
2τ
+ U Kξ(3) (Iξ U, U ) + U Kξ(3)(U, Iξ U ) + U ∂ξ ∂η−1 Kη(2) (Iξ U, U ) +
+ U ∂ξ ∂η−1 Kη(2) (U, Iξ U ) + 2(Iη U )Kη(2) + 2(Iη U )∂ξ ∂η−1 Kη(2) +
+ U ∂η ∂ξ−1 Kξ(3) (Iη U, U ) + U ∂η ∂ξ−1 Kξ(3) (U, Iη U ) +
1
+ U Kη(2) (Iη U, U ) + U Kη(2)(U, Iη U ) − 2 2(Iξ Iη U )Kξ(3) +
2τ

+ 2 Iξ2 U ∂ξ ∂η−1 Kη(2) + (Iη U )Kξ(3)(Iξ U, U ) + (Iη U )Kξ(3) (U, Iξ U ) +
+ (Iξ U )∂ξ ∂η−1 Kη(2)(Iξ U, U ) + (Iξ U )∂ξ ∂η−1 Kη(2) (U, Iξ U ) +
+ 2(Iξ2 U )Kη(3) + 2(Iη Iξ U )∂ξ ∂η−1 Kη(2) + (Iη U )∂η ∂ξ−1 Kξ(3) (Iη U, U ) +
+ (Iη U )∂η ∂ξ−1 Kξ(3) (U, Iη U ) +

+ (Iξ U )Kη(2)(Iη U, Ū ) + (Iξ U )Kη(2)(U, Iη U ) ,
R1 = Iξ R0 + 2τ1 2 Qξ W, R2 = Iη R0 + 2τ1 2 Qη W . In view of inequalities (4.2) we

obtain the following estimate kR0 k 6 Cε 3 τ 3γ −2 and kRj kL∞ 2 + kR2j kL∞ L2 6
ξ Lη ξ η
3 3γ −2
Cε τ , j = 0, 1 for the remainder terms. Taking into account the commutator
relation [L̃, Ij ] = 0, j = ξ, η, where L̃ = i∂τ + 4τ1 2 (∂ξ2 + ∂η2 ), we get from
Equation (4.8)
1
L̃Wj + QWj + Rj = 0, j = 0, 1, 2, (4.13)
τ
for the functions W0 = W, W1 = Iξ W and W2 = Iη W . From (4.13) we eas-
ily get the estimate kW0 k < ε. In order to prove the estimates kWj kL∞ 2 +
η Lξ
kW2j kL∞ 2 < 20ε, j = 0, 1, we define the following evolution operator Yξ (τ )ϕ =
ξ Lη √ R
F −1 eip /4τ ϕ̂(p, q) = √iπτ e−iτ (ξ −x) ϕ(x, y) dx. Operator Yη (τ ) is defined anal-
2 2
ogously. Operators Yξ and Yη are unitary in L2 so we have kYξ (τ )φk √ = kφk =

kYη (τ )φk. Also, it is easy to see that the estimates kYξ (τ )φkL∞ 2
ξ Lη
6 τ kφkL1ξ L2η ,
−1)φ̂(p, q)kL1p L2q 6 τ −α/2 (kφk+kφξ k) and simi-
2 /4τ
k(Yξ (τ )−1)φkL∞ 2 6 k(e
ξ Lη
ip
larly k(Yξ (τ ) − 1)φk 6 τ −α (kφk + kφξ k) are valid, where α ∈ [0, 12 ). Multiplying
(4.13) by Yξ (−τ ) and Yη (−τ ) we obtain
1
L̃η Yξ (−τ )Wj + Q(Yξ (−τ )Wj ) + A2j +1 = 0,
τ
1
L̃ξ Yη (−τ )W2j + Q(Yη (−τ )W2j ) + A2j +2 = 0, j = 0, 1, (4.14)
τ
where L̃j = i∂τ + 1 2
∂ ,
4τ 2 j
j = ξ, η, the remainder terms are

j V
A2j +1 = Yξ (−τ )Rj + (Yξ (−τ ) − 1) 1 − Iξ2 − Iη2 τ Iξ N −
τ
1
− Q(Yξ (−τ ) − 1)Wj ,
τ
j = 0, 1, and similarly

j V
A2j +2 = Yη (−τ )Rj + (Yη (−τ ) − 1) 1 − Iξ2 − Iη2 τ Iη N −
τ
1
− Q(Yτ (−τ ) − 1)W2j ,
τ
3 3γ −3/2
j = 0, 1. Via (4.2) we have kA2j +1 kL∞ 2 +kA2j +2 kL∞ L2 6 Cε τ
ξ Lη
, j = 0, 1.
η ξ
√
Integrating (4.14) with respect to time τ and using the representations τ kwkL∞ 2
x Ly
= kW kL∞ 2
ξ Lη
6 kY ξ (−τ )W k ∞
Lξ Lη2 + k(1 − Y ξ (−τ ))W k ∞ 2
Lξ Lη 6 kW (1)k ∞ 2
Lξ Lη +
10ετ −δ , δ = 14 −3γ > 0 by virtue of (4.2) we obtain the inequality supt ∈[0,T ] (kw(t)
kL∞ 2 + kwx (t)kL∞ L2 + kw(t)kL∞ L2 + kwy (t)kL∞ L2 ) < 100ε. The contradiction
x Ly x y y x y x
obtained proves estimate (4.1) for all t > 0. Now a standard continuation argument
yields the result of Theorem 1.3. 2
The proof of Theorem 1.4 is simpler than that of Theorem 1.3, so we leave it to
the reader.
5. Large Time Behavior
Proof of Theorem 1.5. Now we assume that the initial data decay more rapidly at
infinity u0 ∈ H4,1 ∩ H3,2 . So in the same way as in the proof of Theorem 1.3 we
can obtain the following estimate
(1 + t)−γ (kωk2,1 + kU(−t)ωk1,2 ) + (1 + t)kωk∞ < 100ε. (5.1)
Analogously to (4.14) we obtain

1
(Y(−τ )W )τ + Q(Y(−τ )W ) + A = 0, (5.2)
τ
where the remainder term equals
V
A = Y(−τ )R0 + (Y(−τ ) − 1)(1 − Iξ2 − Iη2 )τ N −
τ
1
− Q(Y(−τ ) − 1)W.
τ
Via (5.1) we have kAkp 6 Cε 3 τ 3γ −5/4 , p = 2, ∞. So, changing the dependent
variable as follows
Z τ
dr
Y(−τ )W (τ, ξ, η) = g(τ, ξ, η) exp i Q(r, ξ, η)
1 r
in Equation (5.2), we get
Z τ
dr
i∂τ g(τ ) + exp − i Q(r, ξ, η) A = 0. (5.3)
1 r
Integration of (5.3) with respect to time τ yields
Z τ
kg(τ ) − g(s)kp 6 Cε 3 r 3γ −3/2 dr 6 Cε 3 s −δ , (5.4)
s
for all 1 < s < τ , where p = 2, ∞ and δ = 14 − 3γ > 0. Therefore there exist
unique limits W + ∈ L∞ ∩ L2 such that limt →∞ g(t) = W + in L∞ ∩ L2 . Hence
there exists a unique limit Q+ = limτ →∞ Q(τ ) in L∞ . Thus we get
(1 − 1)u(t, x, y)
1
= eiτ ξ +iτ η W
2 2
τ
1 1
= eiτ ξ +iτ η Y(−τ )W + eiτ ξ +iτ η (1 − Y(−τ ))W
2 2 2 2
τ τ Z τ
1 iτ ξ 2 +iτ η2 + x y x y ds
= e W τ, , exp i Q s, , + O t −1−δ
τ 2τ 2τ 1 2τ 2τ s
(5.5)
uniformly with respect to (x, y) ∈ R2 , here δ = 14 − 3γ > 0. For the phase of

Rτ
the asymptotic representation (5.5) we write the identity 1 Q(s) dss = Q+ log τ +
Rτ
8(τ ), where 8(τ ) = (Q(τ ) − Q+ ) log τ + 1 (Q(s) − Q(τ )) dss . We have
Z τ
dr τ
8(τ ) − 8(s) = (Q(r) − Q(τ )) + (Q(τ ) − Q+ ) log , (5.6)
s r s
for all 1 < s < τ . Applying estimates (5.1) and (5.4) to (5.6) we get k8(τ ) −
8(s)k∞ 6 Cεs −δ for 1 < s < τ . This implies that there exists a unique limit
8+ = limτ →∞ 8(τ ) ∈ L∞ such that k8(τ ) − 8+ k∞ 6 Cτ −δ , whence we find
Z τ

x y ds +

Q s, , − Q log τ − 8 6 Cετ −δ .
+
(5.7)
1 2τ 2τ s ∞
We now put W + = W + exp(i8+ ). Therefore we obtain the asymptotics (1.10) for

t → ∞ uniformly with respect to (x, y) ∈ R2 . Teorem 1.5 is proved. 2
Theorem 1.6 is proved in the same way as in the proof of Theorem 1.5, and so
we omit it.
References
1. Ablowitz, J. M. and Haberman, R.: Nonlinear evolution equations in two and three dimensions,
Phys. Rev. Lett. 35 (1975), 1185–1188.
2. Ablowitz, J. M. and Fokas, A. S.: On the inverse scattering transform of multidimensional
nonlinear equations, J. Math. Phys. 25 (1984), 2494–2505.
3. Ablowitz, J. M. and Segur, A.: Solitons and Inverse Scattering Transform, PA, SIAM, 1981.
4. Anker, D. and Freeman, N. C.: On the soliton solutions of the Davey–Stewartson equation for
long waves, Proc. R. Soc. A 360 (1978), 529–540.
5. Bardos, C., Sulem, C. and Sulem, P. L.: On the continuous limit for a system of classical spins,
Comm. Math. Phys. 107 (1986), 431–454.
6. Benney, D. J. and Roskes, G. L.: Wave instabilities, Stud. Appl. Math. 48 (1969), 377–385.
7. Beals, R. and Coifman, R. R.: The spectral problem for the Davey–Stewartson and Ishimori
hierarchies, Proc. Conf. on Nonlinear Evolution Equations: Integrability and Spectral Methods,
Manchester, U.K., 1988.
8. Chihara, H.: The initial value problem for the elliptic–hyperbolic Davey–Stewartson equation,
Preprint, 1995.
9. Constantin, P.: Decay estimates of Schrödinger equations, Comm. Math. Phys. 127 (1990),
101–108.
10. Cornille, H.: Solutions of the generalized nonlinear Schrödinger equation in two spatial
dimensions, J. Math. Phys. 20 (1979), 199–209.
11. Davey, A. and Stewartson, K.: On three-dimensional packets of surface waves, Proc. R. Soc. A
338 (1974), 101–110.
12. Djordjevic, V. D. and Redekopp, L. G.: On two-dimensional packets of capillary-gravity waves,
J. Fluid Mech. 79 (1977), 703–714.
13. Doi, S.: On the Cauchy problem Schrödinger type equations and regularity of solutions, J.
Math. Kyoto Univ. 34 (1994), 319–328.
14. Fokas, A. S. and Santini, P. M.: Recursion operators and bi-Hamiltonian structures in
multidimensions. I, II, Comm. Math. Phys. 115 (1988), 375–419, 449–474.
15. Fokas, A. S. and Sung, L. Y.: On the solvability of the N-wave, Davey–Stewartson and
Kadomtsev–Petviashvili equations, Inverse Problems 8 (1992), 673–708.
16. Ghidaglia, J. M. and Saut, J. C.: On the initial value problem for the Davey–Stewartson systems,
Nonlinearity 3 (1990), 475–506.
17. Hayashi, N.: Local existence in time of small solutions to the Davey–Stewartson system,
Annales de l’I.H.P. Physique Theorique 65 (1996), 313–366.
18. Hayashi, N.: On the initial value problem for the Davey–Stewartson and the Ishimori sys-
tems, in The Proceedings of the 4th MSJ International Reserch Institute on “Nonlinear
Waves”, Sapporo, July 10–21, 1995, GAKUTO International Series, Mathematical Sciences
and Applications, Gakkotosho, vol. 10, 1997, pp. 145–154.
19. Hayashi, N.: Local existence in time of solutions to the elliptic–hyperbolic Davey–Stewartson
system without smallness condition on the data, J. Analysé Mathématique 73 (1997), 133–164.
20. Hayashi, N.: Local existence in time of small solutions to the Ishimori system, Diff. Integral
Eqs., to appear.
21. Hayashi, N. and Hirata, H.: Global existence and asymptotic behavior in time of small solutions
to the elliptic–hyperbolic Davey–Stewartson system, Nonlinearity 9 (1996), 1387–1409.
22. Hayashi, N. and Hirata, H.: Local existence in time of small solutions to the elliptic–
hyperbolic Davey–Stewartson system in the usual Sobolev space, Proceedings of the Edinburgh
Mathematical Society 40 (1997), 563–581.
23. Hayashi, N. and Naumkin, P. I.: Asymptotic behavior in time of solutions to the derivative non-
linear Schrödinger equation revisited, Discrete and Continuous Dynamical Systems 3 (1997),
383–400.
24. Hayashi, N. and Saut, J. C.: Global existence of small solutions to the Davey–Stewartson and
the Ishimori systems, Diff. Integral Eqs. 8 (1995), 1657–1675.
25. Hayashi, N. and Saut, J. C.: Global existence of small solutions to the Ishimori system without
exponential decay of the data, Diff. Integral Eqs. 9 (1996), 1183–1195.
26. Ishimori, Y.: Multi vortex solutions of a two dimensional nonlinear wave equation, Progr.
Theor. Phys. 72 (1984), 33–37.
27. Kato, T.: On the Cauchy problem for the (generalized) Korteweg–de Vries equation, Advances
in Mathematics Supplementary Studies, Studies in Appl. Math. 8 (1983), 93–128.
28. Kenig, C. E., Ponce, G. and Vega, L.: Well-posedness and scattering results for the generalized
Korteweg–de Vries equation via contraction principle, Comm. Pure Appl. Math. 46 (1993),
527–620.
29. Konopelchenko, B. G. and Matkarimov, B. T.: Inverse spectral transform for the nonlinear
evolution equation generating the Davey–Stewartson and Ishimori equations, Stud. Appl. Math.
82 (1990), 319–359.
30. Linares, F. and Ponce, G.: On the Davey–Stewartson systems, Ann. Inst. Henri Poincaré, Anal.
non linéaire 10 (1993), 523–548.
31. Soyeur, A.: The Cauchy problem for the Ishimori equations, J. Funct. Anal. 105 (1992), 233–
255.
32. Stein, E. M.: Singular Integral and Differentiability Properties of Functions, Princeton Univ.
Press, Princeton Math. Series 30, 1970.
33. Sung, L. Y.: An inverse-scattering transform for the Davey–Stewartson II equations, Part III,
J. Math. Anal. Appl. 183 (1994), 477–494.
34. Sung, L. Y.: The Cauchy problem for the Ishimori equation, Preprint, 1993.
83
Stochastic Isometries in Quantum Mechanics
P. BUSCH
Department of Mathematics, University of Hull, HU6 7RX Hull, U.K.
e-mail: p.busch@maths.hull.ac.uk
(Received: 3 December 1998, accepted: 31 March 1999)

Abstract. The class of stochastic maps, that is, linear, trace-preserving, positive maps between the
self-adjoint trace class operators of complex separable Hilbert spaces plays an important role in the
representation of reversible dynamics and symmetry transformations. Here a characterization of the
isometric stochastic maps is given and possible physical applications are indicated.
Mathematics Subject Classifications (1991): Primary: 47D45; Secondary: 47D20, 81Q99, 81R99.
Key words: Hilbert space, trace class, state operator, stochastic map, isometry, quantum mechanics,
reversibility.
1. Introduction
The mathematical modeling of a quantum dynamical system is based on the dual
concepts of states and observables. Of particular importance for the representation
of dynamics and symmetries are structure-preserving bijective maps of the sets
of states and observables, respectively. Such maps, referred to as affine automor-
phisms of states and Jordan automorphisms of observables, respectively, have been
quite thoroughly studied in the Hilbert space model ([1], for a recent systematic
account, see [4]) and in the C ∗ -algebraic formulation of quantum theory [7]. These
automorphisms are special instances of the larger classes of linear isometries act-
ing on the state space and the algebra of observables, respectively, which are of
interest in their own right. While isometries of operator algebras were analyzed by
Kadison in 1951 [8], applications of isometric transformations of states have been
considered only in recent years. The characterization of the class of isometric state
transformations in the Hilbert space model of quantum mechanics is the subject of
the present paper.
The paper is organized as follows. In Section 2 the notion of a linear state trans-
formation – also called stochastic map – in the Hilbert space model of quantum
mechanics is presented and some of its basic properties are reviewed. An isometric
map among the stochastic maps is distinguished by a feature that turns out to be
fundamental for the subsequent considerations: it sends orthogonal pairs of states to
orthogonal pairs (Proposition 2.1). In Section 3 the main result (Theorem 1) con-
cerning isometric stochastic maps is stated: any stochastic isometry decomposes
84 P. BUSCH
into a convex combination of pure stochastic isometries onto mutually orthogonal

ranges. In the case of a completely positive stochastic isometry, all these pure
isometries are induced by unitary maps (Theorem 2). The decomposition of sto-
chastic isometries involves two steps (Propositions 3.3 and 3.4), proofs of which
are carried out in Sections 4 and 5. In the first step one is led to introduce the
concept of mixing isometries (Definition 2), while the second step consists of an
analysis of the structure of mixing isometries.
Section 6 concludes with some general observations and an outline of physical
applications. First, a stochastic isometry is shown to allow an interpretation as a
model of a reversible state transformation even when surjectivity is not given (The-
orem 3). An isometric state transformation leads to a reduction of the state space
so that some observables and symmetries can no longer be distinguished. This may
be taken as a reversible model of structure formation. Finally, a stochastic isometry
can be used in conjunction with a certain transformation of observables to produce
equivalent descriptions of the same physical system; these occur naturally in the
construction of extended models accounting for new experiences with the given
system.
2. Stochastic Maps
Let H and H̃ be complex separable Hilbert spaces with inner products h· | ·i, taken
to be linear in the second argument and conjugate linear in the first. Let B∞ (H)
be the complex Banach space of linear operators of trace class on H, with trace
tr[·] and trace norm k · k1 . Its dual space can be identified with the Banach space
B(H) of bounded linear operators on H. The self-adjoint parts of B∞ (H) and
B(H), V := B1 (H)s and W := B(H)s , are real Banach spaces. By V + we
denote the set of positive linear operators in V , the convex positive cone of V .
With reference to the quantum physical application, V is called state space, the
elements of the subset S := V + ∩ {ρ ∈ V : tr[ρ] = 1} are called states. The
set S is convex, its extreme points are given by the orthogonal projections of rank
1. These are called pure states and denoted Pϕ , ϕ ∈ H, ϕ 6= 0. For H̃, we
denote the corresponding entities as Ṽ , W̃ , S̃, etc. A linear map T : V → Ṽ is
called a stochastic map (on V ) if it is positive [T (V + ) ⊂ Ṽ + ] and trace preserving
[tr ◦ T = tr], or equivalently, if it sends states to states [T (S) ⊆ S̃]. A stochastic
map that is isometric [kT (ρ)k1 = kρk1 , ρ ∈ V ] will be called stochastic isometry.
The term operator will be used throughout to denote a linear operator. By 0 and
I (or IH ) we will denote the zero and identity operators, respectively. According
to the spectral theorem, any self-adjoint operator a admits a decomposition into
positive and negative parts, denoted a = a+ − a− , where the ranges of a+ , a−
are mutually orthogonal subspaces. For a bounded operator a we define |a| :=
(a ∗ a)1/2 . If a is self-adjoint, one has |a| = a+ + a− . For ρ ∈ V we have kρk1 =
tr[ρ+ ] + tr[ρ− ] = kρ+ + ρ− k1 , and so
ρ ∈ V + ⇐⇒ kρk1 = tr[ρ]. (2.1)
STOCHASTIC ISOMETRIES IN QUANTUM MECHANICS 85
We recall some basic facts about stochastic maps and stochastic isometries.
A linear map T : V → Ṽ is called contractive [or a contraction], if for all
ρ ∈ V , kT (ρ)k1 6 kρk1 .
LEMMA 2.1. Let T : V → Ṽ be a linear map. The following are equivalent:

(i) T is trace preserving and positive (i.e., a stochastic map);
(ii) T is trace preserving and contractive;
(iii) T (S) ⊆ S̃.
The proof is a straightforward application of Equation (2.1).
DEFINITION 1. Positive elements ρ, σ ∈ V + \{0} are called orthogonal, ρ ⊥ σ ,

if ρ · σ = 0. A stochastic map T : V → Ṽ is said to be orthogonality preserving if
ρ ⊥ σ implies T (ρ) ⊥ T (σ ) for all ρ, σ ∈ S [and thus for all ρ, σ ∈ V + \{0}].
By the spectral theorem, ρ ⊥ σ exactly when ρ and σ are the positive and
negative parts of ρ − σ , respectively. Hence we have:
ρ ⊥ σ ⇐⇒ kρ − σ k1 = kρ + σ k1 , ρ, σ ∈ V + \{0}. (2.2)
PROPOSITION 2.1. Let T : V → Ṽ be a stochastic map. The following are

equivalent:
(i) T is an isometry;
(ii) T is orthogonality preserving;
(iii) T is orthogonality preserving for pairs of pure states.
The proof is a straightforward application of Equation (2.2).
3. Stochastic Isometries
In this section the main results concerning stochastic isometries will be stated. The
proofs will be developed in Sections 4 and 5.
A stochastic map is called pure if it sends pure states to pure states. An iso-
metric linear or antilinear map U : H → H̃ will be called unitary or antiunitary,
respectively. We first present the known case of surjective stochastic isometries.
PROPOSITION 3.1. Let T : V → Ṽ be a surjective stochastic isometry. Then T

is of the form
T (ρ) = UρU ∗, ρ ∈ V, (3.1)
where U : H → H̃ is unitary or antiunitary.

Proof. Noting that the inverse of a surjective stochastic isometry T is a trace
preserving isometric map on V and hence, by Lemma 2.1, a stochastic isometry,
it is easy to verify that T is pure. By Theorem 2.3.1 of [5] it follows that T is
86 P. BUSCH
of the form T (ρ) = UρU ∗ , ρ ∈ V , where U is linear or antilinear. Furthermore,

U ∗ U = IH since T is trace-preserving, and U U ∗ = IH̃ by surjectivity of T . Hence
U is unitary or antiunitary. 2
If H = H̃ is finite-dimensional then any stochastic isometry T : V → V is

surjective, hence pure and of the form (3.1).
An affine automorphism of the convex set of states S is a bijective affine map of
S onto itself. Such maps are taken to represent symmetries on the set of states. Any
affine map on S has a unique extension to a linear map on V , which is a stochastic
map. An affine automorphism of S extends to a surjective stochastic isometry. Thus
it is seen that in the case H = H̃, Proposition 3.1 reproduces the Wigner–Kadison
characterization of symmetries [1, 4, 7].
PROPOSITION 3.2. Let H, H̃ be complex separable Hilbert spaces such that H̃
L be presented as a direct sum of mutually orthogonal closed subspaces, H̃ =
can
( N k=1 H̃k ) ⊕ H̃0 , N ∈ N ∪ {∞}, with dim H̃k = dim H, k = 1, 2,P . . . , N. Let
Uk : H → H̃k be unitary or antiunitary maps, w1 , w2 , . . . ∈ (0, 1), k wk = 1.
Then
X
N
T (ρ) = wk Uk ρUk∗ , ρ ∈ V, (3.2)
k=1
defines a stochastic isometry T : V → Ṽ . In the case N = ∞, the sum converges

in trace norm. P ∗
Proof. The trace is a normal map and so tr[T (ρ)] = N k=1 wk tr[Uk Uk ρ] = tr[ρ]
∗
for ρ ∈ V , the last equality being due to Uk Uk = I . Hence T is trace preserving.
T is positive since Uk ρUk∗ ∈ V + for ρ ∈ V + . All Tk : ρ 7→ Uk ρUk∗ are isometric,
P
and for ρ ∈ V , |Tk (ρ)| ⊥ |T` (ρ)| if k 6= `. Thus, |T (ρ)| = N k=1 wk |Tk (ρ)|, and
X
N X
N
kT (ρ)k1 = wk kTk (ρ)k1 = wk kρk1 = kρk1 .
k=1 k=1
Pn+m Pn+m
Convergence in trace norm follows from k k=n wk Uk ρUk∗ k1 6 k=n wk kρk1
and the fact that Ṽ is complete. 2
The construction of Proposition 3.2 turns out to be generic.

THEOREM 1. Let L T : V → Ṽ be a stochastic isometry. Then H̃ can be decom-
posed as H̃ = ( N k=1 H̃k ) ⊕ H̃0 ,P
with dim(H̃k ) = dim(H), k = 1, 2, . . . , N 6
N
∞. There exist weights wk > 0, k=1 wk = 1, and unitary or antiunitary maps
Uk : H → H̃k such that T is of the form of Equation (3.2).
Theorem 1 is an immediate consequence of Propositions 3.3 and 3.4 to be for-
mulated next. A concept central to the decomposition of stochastic isometries is
the following.
DEFINITION 2. A stochastic map T : V → Ṽ will be called (m)-mixing (m ∈

N) if every pure state Pϕ , ϕ ∈ H\{0}, is sent to an m-fold degenerate mixture; that
is: T (Pϕ ) = m1 5ϕ , where 5ϕ is an orthogonal projection of rank m.
PROPOSITION 3.3. A stochastic isometry T : V → Ṽ can be decomposed

into a (σ -)convex combination of mixing stochastic isometries with mutually or-
thogonal ranges. That is, there is a family of mixing stochastic isometries Tν :
V → Ṽ , ν = 1,P2, . . . , N 6 ∞ and a strictly decreasing sequence of weights
w1 > w2 > · · · , N ν=1 wν = 1, such that
X
N
T (ρ) = wν Tν (ρ), ρ ∈ V. (3.3)
ν=1
The sum converges in trace norm when N = ∞. The ranges of the Tν can be rep-
resented as subspaces of B1 (H̃ν )s , where the H̃ν are mutually orthogonal closed
subspaces of H̃.
PROPOSITION 3.4. Any m-mixing stochastic isometry T : V → Ṽ is of the form

given in Equation (3.2), with N = m, wk = m1 .
A positive map T : B1 (H)s → B1 (H̃)s is called completely positive if its

canonical extension T (n) : B1 (H)s ⊗ B1 (Cn )s → B1 (H̃)s ⊗ B1 (Cn )s is positive
for all n ∈ N. As a specification of Theorem 1 the following characterization of
completely positive stochastic isometries will be obtained.
THEOREM 2. A completely positive stochastic isometry T : V → Ṽ is of the
form of Equation (3.2), with all Uk unitary, k = 1, 2, . . . , N.
4. Decompositions of Stochastic Isometries

For ρ ∈ B1 (H), a ∈ B(H), we will interchangeably use the notations hρ, ai
and ha, ρi for tr[ρ · a]. The dual T ∗ : B(H̃)s → B(H)s of a positive linear map
T : B1 (H)s → B1 (H̃)s is a normal positive linear map ([5, Lemma 2.2.2]), and
any normal positive linear map on B(H̃)s is the dual of a positive linear map on
B1 (H)s . T ∗ (I ) = I is equivalent to T being trace-preserving. Maps T on B1 (H)s
and their dual maps T ∗ will henceforth be understood to be linear without explicit
mention.
DEFINITION 3. The support (projection) of a state ρ ∈ S, denoted 5(ρ), is the
smallest orthogonal projection P such that hρ, P i = 1.
In the sequel the term projection will mean orthogonal projection. We shall
freely use properties of the complete orthocomplemented lattice of projections.
Projections P , Q are called orthogonal, P ⊥Q, if P Q = 0, or equivalently, P +
88 P. BUSCH
⊥
Q 6 I , or P 6 I − Q =: W Q . The supremum of a family of projections
P , α ∈ A, will be denoted α∈A P (α) . Note that the projections 5ϕ introduced
(α)
in Definition 2 are the support projections of T (Pϕ ).
LEMMA 4.1. Let T : V → Ṽ be a stochastic map. The following are equiva-

lent:
(i) T is isometric;
(ii) T ∗ (5(T (ρ))) = 5(ρ) for all ρ ∈ S;
(iii) T ∗ (5(T (Pϕ ))) = Pϕ for all ϕ ∈ H\{0}.
Proof. Assume that T is a stochastic isometry. Let ρ ∈ S, then 0 6 T ∗ (5(T (ρ)))

6 I , and hρ, T ∗ (5(T (ρ)))i = hT (ρ), 5(T (ρ))i = 1. Therefore, 5(ρ) 6 T ∗ (5(T
(ρ))). We show that P also T ∗ (5(T (ρ))) 6 5(ρ), so that both operators are equal.
Write 5(ρ)⊥ = k Pϕk . Then Pϕk ⊥ 5(ρ), and so Pϕk ⊥ ρ. Since T is an or-
thogonality preserving stochastic map [by Proposition 2.1], it follows that T (Pϕk ) ⊥
T (ρ). Then
0 = hT (Pϕk ), 5(T (ρ))i = hPϕk , T ∗ (5(T (ρ)))i.
This gives Pϕk ·T ∗ (5(T (ρ))) = 0 and therefore one has 5(ρ)⊥ ·T ∗ (5(T (ρ))) = 0.
Since T ∗ (5(T (ρ))) 6 I , it follows that T ∗ (5(T (ρ))) 6 5(ρ). Thus, T ∗ (5(T (ρ)))
= 5(ρ).
Conversely, assume (ii) holds. We show that the stochastic map T is orthogo-
nality preserving and thus, by Proposition 2.1, an isometry. Let ρ, σ ∈ S, ρ ⊥ σ .
Then hT (ρ), 5(T (σ ))i = hρ, T ∗ (5(T (σ )))i = hρ, 5(σ )i = 0. This implies
T (ρ) ⊥ T (σ ).
Property (iii) is entailed by (ii). The converse implication is a fairly straightfor-
ward consequence of the fact that T ∗ is normal. 2
LEMMA 4.2. Let T : V → Ṽ be a stochastic isometry and not pure. There exists
a strictly decreasing sequence w̃ν , 0 <Pw̃ν < 1, ν = 1, 2 . . . , N 6 ∞, and a
sequence of numbers mν ∈ N such that ν w̃ν mν = 1 and
X
N
T (Pϕ ) = w̃ν 5νϕ (4.1)
ν=1
for all ϕ ∈ H\{0}. Here the 5νϕ are projections with rank mν . Furthermore,
µ
5νϕ ⊥ 5ψ for all ϕ, ψ ∈ H\{0} if ν 6= µ.
Proof. For ϕ, ψ 6= 0, write the spectral decompositions of T (Pϕ ), T (Pψ ) as
X
N X
M
T (Pϕ ) = aν Pϕν , T (Pψ ) = bµ Pψµ , (4.2)
ν=1 µ=1
respectively, where the ϕν and ψµ form orthonormal systems, N, M ∈ N ∪ {∞},

and the aν , bµ are the nonzero eigenvalues. We show that the eigenvalues coincide
in numerical values and multiplicities. First observe that Pϕν 6 5(T (Pϕ )), so that,
by virtue of Lemma 4.2, T ∗ (Pϕν ) 6 Pϕ , hPϕ , T ∗ (Pϕν )i = hT (Pϕ ), Pϕν i = aν , and
therefore
T ∗ (Pϕν ) = aν Pϕ . (4.3)
Similarly,
T ∗ (Pψµ ) = bµ Pψ . (4.4)
Case 1: 0 =6 |hϕ | ψi|(6= 1). Take unit vectors ϕ 0 , ψ 0 in the span of ϕ, ψ such that
hϕ | ϕ i = hψ | ψ 0 i = 0. Then Pϕ + Pϕ0 = Pψ + Pψ 0 and therefore T (Pϕ ) + T
0
(Pϕ0 ) = T (Pψ ) + T (Pψ 0 ). Since T is orthogonality preserving, this yields:

2
T (Pϕ ) = T (Pψ )T (Pϕ ) + T (Pψ 0 )T (Pϕ ),
and so
2 2
T (Pψ ) T (Pϕ ) = T (Pψ ) T (Pϕ ),
and hence by induction

n m m n
T (Pψ ) T (Pϕ ) = T (Pψ ) T (Pϕ ) , n, m ∈ N. (4.5)
On multiplication of Equation (4.5) with Pψµ from the left and with Pϕν from the
right, we obtain
bµn aνm Pψµ Pϕν = bµm aνn Pψµ Pϕν . (4.6)
Choosing n = 2, m = 1, one concludes that
aν = bµ whenever Pψµ Pϕν 6= 0. (4.7)
Applying (4.5) again, one can write

n−1
n−1
T (Pψ ) , T (Pϕ ) = T (Pψ ), T (Pϕ ) , n > 2,
that is, by virtue of (4.2),

X X
bµn−1 hPψµ , T (Pϕ )i = aνn−1 hT (Pψ ), Pϕν i.
µ ν
Observing that (by virtue of (4.3), (4.4))
hPψµ , T (Pϕ )i = hT ∗ (Pψµ ), Pϕ i = bµ hPψ , Pϕ i, (4.8)
hT (Pψ ), Pϕν i = hPψ , T ∗ (Pϕν )i = aν hPψ , Pϕ i, (4.9)

90 P. BUSCH
and using the fact that hPψ , Pϕ i 6= 0, we obtain

X X
aνn = bµn , for all n ∈ N. (4.10)
ν µ
We rewrite this, making the multiplicities explicit: thus we let a1 , a2 , . . . and b1 , b2 ,

. . . denote the strictly decreasing sequences of eigenvalues of T (Pϕ ), T (Pψ ), with
multiplicities nν , mµ , respectively. Then Equation (4.10) reads:
X X
nν aνn = mµ bµn , for all n ∈ N. (4.11)
ν µ
The orthonormal systems of eigenvectors shall now be denoted ϕν,i , ψµ,j , where
i ∈ {1, . . . , nν }, j ∈ {1, . . . , mµ }. Now observe that due to (4.8), (4.9),
hT (Pψ ), Pϕν,i i = hT (Pψ ), T ∗ (Pϕν,i )i = aν hPψ , Pϕ i,
which implies that for each ν there must be µ, j such that hPψµ,j , Pϕν,i i 6= 0.
Therefore, by (4.7), aν = bµ . A similar reasoning entails that for each µ there
must be ν such that aν = bµ . Since the two sequences of eigenvalues are strictly
decreasing, they must be identical: aν = bν , for all values of ν. It remains to be
shown that the multiplicities coincide as well. Equation (4.11) can be written as
X
N
(nν − mν )aνn = 0, for all n ∈ N. (4.12)
ν=1
We show that xν := nν − mν = 0 for all ν = 1, . . . , N. This is obvious for N = 1.

So let N > 1. Suppose x1 6= 0. Then (4.12) gives
X
N
0 =1+ (a1n x1 )−1 aκn xκ ,
κ=2
that is,
N
X aκ n xκ
= −1.
κ=2
a1 x1
But then
X
N aκ n xκ X aκ n xκ
N

1= 6 =: S(n) (< ∞).
aκ=2
x 1 a 1 x κ=2 1 1
Since a1 > a2 > · · · , we also have:
aκ n+1 xκ a2 n
XN
S(n + 1) = 6 S(1).
κ=2
a1 x1 a1
Hence S(n) ↓ 0 and S(n) > 1, which is a contradiction. Therefore, x1 = 0. The

argument can be repeated for x2 , x3 , . . . to yield xν = 0 for all ν.
Note that Equation (4.7) also implies
hPψµ,j , Pϕν,i i = 0 for ν 6= µ. (4.13)
Finally we get, putting aν = w̃ν :
! !
X
N X
mν X
N X
mν
T (Pϕ ) = w̃ν Pϕν,i , T (Pψ ) = w̃ν Pψν,j , (4.14)
ν=1 i=1 ν=1 j =1
P
where ν w̃ν nν = 1. This is of the form of Equation (4.1), and due to (4.13) the
projections 5νϕ , 5µψ are mutually orthogonal for any ϕ, ψ 6= 0 (hϕ|ψi 6= 0) if
ν 6= µ.
Case 2: hϕ|ψi = 0. Replace ψ by a unit vector ψ̃ in the span of ϕ, ψ and not
parallel to ϕ or ψ. Then apply the argument of Case 1 to obtain Equations (4.14)
for the pairs ϕ, ψ̃ and ψ̃, ψ. It follows that equations of this form also hold for
orthogonal pairs ϕ, ψ. 2
LEMMA 4.3. Let T : V → Ṽ be a stochastic isometry. There exists P a complete

family of mutually orthogonal projections Pν , ν = 0, 1, . . . , N 6 ∞, Nν=0 Pν =
I , such that
X
N
T (ρ) = Pν T (ρ)Pν . (4.15)
ν=0
For ν = 1, . . . , N, ϕ ∈ H\{0}, let 5νϕ be the projections of rank mν established

W P
in Lemma 4.2. Then Pν = ϕ∈H\{0} 5νϕ for ν = 1, . . . , N, and P0 = I − N ν=1 Pν .
Proof. The projections Pν are mutually orthogonal by virtue ofPthe orthogonality
of all 5νϕ , 5µψ , ν 6= µ. Then Pν 5µϕ = δµν 5νϕ = 5µϕ Pν , and so ν Pν T (Pϕ )Pν =
P
ν w̃ν 5ϕ = T (Pϕ ). By continuity of T , this equality extends to all T (ρ), ρ ∈ V . 2
ν
Proof of Proposition 3.3. For the w̃ν , mν given in Lemma 4.2, put wν = w̃ν mν ,
and define maps Tν via
1 ν
Tν (Pϕ ) := 5 = wν−1 Pν T (Pϕ )Pν ,
mν ϕ
with the projections Pν , ν = 1, 2, . . . , N, given in Lemma 4.3. Then (4.1) realizes
(3.3), and in view of (4.15), the maps Tν are mixing stochastic maps on V with
ranges in B1 (H̃ν )s , H̃ν = Pν H̃.
5. Mixing Stochastic Isometries

We establish some properties of the m-mixing stochastic isometries which are in-
strumental in proving Proposition 3.4. The proof given here applies Hilbert space
92 P. BUSCH
techniques and emphasizes elementary geometrical aspects, particularly the preser-

vation of orthogonality. In the appendix an alternative proof is presented that is
based on a link with Kadison’s characterization of isometries of operator algebras.
LEMMA 5.1. Let T : V → Ṽ be an m-mixing stochastic isometry, with T (Pφ ) =
1
5 , φ ∈ H\{0}. Let ϕ, ψ ∈ H, ξ ∈ H̃ be unit vectors. Then the following
m φ
hold:
1
Pξ 6 5ϕ H⇒ T ∗ (Pξ ) = Pϕ , (5.1)
m
Pξ 6 5ϕ H⇒ hPξ , 5ψ i = hPϕ , Pψ i, (5.2)
5ϕ 5ψ 5ϕ = hPϕ , Pψ i5ϕ . (5.3)
Proof. Let Pξ 6 5ϕ = mT (Pϕ ). Then by Lemma 4.1, 0 6 T ∗ (Pξ ) 6 T ∗ (5ϕ ) =

Pϕ , and therefore T ∗ (Pξ ) = aPϕ , 0 6 a 6 1. Further, 1 = hPξ , 5ϕ i = hT ∗ (Pξ ),
mPϕ i = a · m, hence a = m1 . This proves (5.1).
Let Pξ 6 5ϕ , then with 5ψ = mT (Pψ ) and (5.1) one obtains
hPξ , 5ψ i = mhT ∗ (Pξ ), Pψ i = hPϕ , Pψ i,
that is, (5.2).
Let P {ϕk | k = 1, . . . , m} ∪ {ϕ`0 | ` ∈ L} be an orthonormal basis of H such that
0
5ϕ = m k=1 Pϕk and thus 5ϕ ϕ` = 0 for all `. Then the only nonvanishing matrix
elements of the (finite rank) operator 5ϕ 5ψ Pϕ are those obtained from the ϕk and
(5.3) is equivalent to
hϕj | 5ψ ϕk i = hPϕ , Pψ iδj k . (5.4)
For j = k one has hϕk | 5ψ ϕk i = hPϕk , 5ψ i, which equals √
hPϕ , Pψ i due to (5.2).
For j 6= k, consider the unit vectors η := (ϕj + αϕk )/ 2, |α| = 1. We have
Pη 6 5ϕ and therefore, by (5.2),
hPη , 5ψ i = hPϕ , Pψ i. (5.5)
Now observe that for any unit vectors ϕ, ψ, aϕ + bψ (a, b ∈ C), the projection
Paϕ+bψ can be written as
Paϕ+bψ = |a|2 Pϕ + |b|2 Pψ + a b̄Aϕψ + ābAψϕ , (5.6)
where Aϕψ is the operator of rank one defined via
Aϕψ ξ = ϕhψ | ξ i, ξ ∈ H. (5.7)
Thus we obtain
1 1 1
Pη = P(ϕj +aϕk )/√2 = Pϕj + Pϕk + {aAϕk ϕj + āAϕj ϕk }. (5.8)
2 2 2
Again by (5.2), we have hPϕj , 5ψ i = hPϕk , 5ψ i = hPϕ , Pψ i. Then combining

(5.8) and (5.5) yields
Re{ahϕj | 5ψ ϕk i} = 0.
Choosing for a the values a = 1, i, one concludes that hϕj | 5ψ ϕk i = 0. Hence (5.4)
follows, and so (5.3) is verified. 2
PROPOSITION 5.1. Let T : V → Ṽ be an m-mixing stochastic map. The follow-

ing statements are equivalent:
(i) T is a stochastic isometry;
(ii) Pξ 6 5ϕ H⇒ T ∗ (Pξ ) = m1 Pϕ for all ϕ ∈ H\{0}, ξ ∈ H̃\{0};
(iii) Pξ 6 5ϕ H⇒ hPξ , 5ψ i = hPϕ , Pψ i for all ξ, ϕ, ψ ∈ H\{0}, ξ ∈ H̃\{0};
(iv) 5ϕ 5ψ 5ϕ = hPϕ , Pψ i5ϕ for all ϕ, ψ ∈ H\{0}.
Proof. According to Lemma 5.1, (i) implies each of the statements (ii), (iii),
(iv). We show that each of the latter statements implies that T is orthogonality
preserving for pairs of pure states, so that, by Proposition 2.1, T is a stochastic
isometry.
Assume (ii) holds. Let Pϕ ⊥Pψ . Take ξ ∈ H̃\{0} such that Pξ 6 5ϕ . Then
hPξ , 5ψ i = hPξ , mT (Pψ )i = hmT ∗ (Pξ ), Pψ i = hPϕ , Pψ i = 0.
Hence Pξ ⊥5ψ . This holds for any Pξ 6 5ϕ , and therefore 5ϕ ⊥5ψ .

Assume (iii) holds. Let Pϕ ⊥Pψ . It follows that for any ξ ∈ H̃\{0} with Pξ 6
5ϕ , hPξ , 5ψ i = hPϕ , Pψ i = 0, and therefore 5ϕ ⊥5ψ .
Assume (iv) holds. Let Pϕ ⊥Pψ . It follows that 5ϕ 5ψ 5ϕ = hPϕ , Pψ i5ϕ = 0,
and so 5ϕ ⊥5ψ . 2
The following generalization of the relation (5.3) will be crucial.
PROPOSITION 5.2. Let T be an m-mixing stochastic isometry. For unit vec-

tors ϕ, ψ, ϑ ∈ H, the following relation holds between the projections 5ϕ =
mT (Pϕ ), 5ψ = mT (Pψ ), 5ϑ = mT (Pϑ ):
5ϕ 5ϑ 5ψ 5ϕ = a5(L)
ϕ + ā5ϕ ,
(A)
a = hϕ | ϑihϑ | ψihψ | ϕi. (5.9)
Here 5(L) (A)

ϕ , 5ϕ are projections independent of ψ, ϑ and determined solely by Pϕ .
ϕ + 5ϕ = 5ϕ .
They satisfy 5(L) (A)
The proof will be based on the study of a number of special cases and on
exploiting the linearity of T . We note two trivial special cases: if any single pair
among the three vectors are mutually orthogonal then the left-hand side of Equation
(5.9) is identically 0; and if any two of these vectors are linearly dependent then
(5.9) reduces to (5.3).
94 P. BUSCH
LEMMA 5.2. Let T be an m-mixing stochastic isometry. For any pair of mutually
orthogonal unit vectors ϕ1 , ϕ2 ∈ H the following holds:
1 + i (L) 1 + i (A)
5ϕ1 5ϕ1 +iϕ2 5ϕ1 +ϕ2 5ϕ1 = 5ϕ1 + 5ϕ1 . (5.10)
4 4
ϕ1 , 5ϕ1 are projections satisfying 5ϕ1 + 5ϕ1 = 5ϕ1 .

Here 5(L) (A) (L) (A)
Proof. We note first that at this stage we do not claim the independence of the
projections 5(L) (A)
ϕ1 , 5ϕ1 of the choice of ϕ2 . This will be established in a later step.
We will frequently use Equations (5.2) and (5.3) without explicit mentioning.
Let {ϕ1k | k = 1, . . . , m} be an arbitrary orthonormal basis (ONB) of the subspace
5ϕ1 H. Then, by virtue of Equation (5.3), for α ∈ C, |α| = 1, the system of vectors
ϕ2k (α) = 25ϕ2 5ϕ1 +αϕ2 ϕ1k , k = 1, . . . , m, form an ONB of 5ϕ2 H. Furthermore,
5ϕ1 +αϕ2 (ϕ1k + ϕ2k (α)) = 5ϕ1 +αϕ2 (ϕ1k ) + 25ϕ1 +αϕ2 5ϕ2 5ϕ1 +αϕ2 (ϕ1k )
= 25ϕ1 +αϕ2 (ϕ1k )
= 2(5ϕ1 + 5ϕ2 )5ϕ1 +αϕ2 (ϕ1k )
= ϕ1k + ϕ2k (α). (5.11)
In the third line we have used the fact that 5ϕ1 +αϕ2 ⊥5ϕ1 −αϕ2 (since T is orthogo-
nality preserving) and that, by virtue of the linearity of T ,
5ϕ1 +αϕ2 + 5ϕ1 −αϕ2 = 5ϕ1 + 5ϕ2 ,
and therefore
5ϕ1 +αϕ2 6 5ϕ1 + 5ϕ2 . (5.12)

0
Let us denote ϕ2k (1) = ϕ2k and ϕ2k (i) = ϕ2k . Then (5.11) entails Pϕ1k +ϕ2k 6
5ϕ1 +ϕ2 , and so (by (5.2))
1 1
hPϕ1k +ϕ2k , 5ϕ1 +iϕ2 i = hPϕ1 +ϕ2 , Pϕ1 +iϕ2 i = |1 + i|2 = . (5.13)
4 2
But from (5.6) we have
1 1 1
Pϕ1k +ϕ2k = Pϕ1k + Pϕ2k + [Aϕ1k ϕ2k + Aϕ2k ϕ1k ].
2 2 2
Using this to evaluate the left-hand side of (5.13), and noting that hPϕ1k , 5ϕ1 +iϕ2 i =
hPϕ2k , 5ϕ1 +iϕ2 i = 12 , we can conclude that
0
Re hϕ2k | 5ϕ1 +iϕ2 ϕ1k i = Re hϕ2k | ϕ2k i = 0. (5.14)
Next we observe that the operator

∗
U := 4 5ϕ2 5ϕ1 +iϕ2 5ϕ1 5ϕ2 5ϕ1 +ϕ2 5ϕ1 = 45ϕ1 5ϕ1 +iϕ2 5ϕ2 5ϕ1 +ϕ2 5ϕ1
is a partial isometry that acts as a unitary operator on 5ϕ1 H; i.e. U ∗ U = U U ∗ =

5ϕ . Take the ONB {ϕ1k : k = 1, . . . , m} to be a set of eigenvectors of this operator,
U ϕ1k = uk ϕ1k , |uk | = 1. It follows that
0
hϕ2k | ϕ2` i = hϕ1k | U ϕ1` i = uk δk` .
Combining this with (5.14), it follows that uk ∈ {+i, −i} for all k. Denote the
spectral projections of the partial isometry U associated with the eigenvalues i, −i
as 5(L) (A)
ϕ1 , 5ϕ1 , respectively. Then the spectral decomposition of U is

U = i 5(L)ϕ1 − 5ϕ1 , 5ϕ1 + 5ϕ1 = 5ϕ1 .
(A) (L) (A)
(5.15)
Finally we can write the operator on the left-hand side of (5.10) as
5ϕ1 5ϕ1 +iϕ2 5ϕ1 +ϕ2 5ϕ1 = 5ϕ1 5ϕ1 +iϕ2 (5ϕ1 + 5ϕ2 )5ϕ1 +ϕ2 5ϕ1
1
= (5ϕ1 + U ),
4
where by virtue of (5.15) the last expression equals the right-hand side of (5.10). 2
LEMMA 5.3. Let dim H > 2 and T : V → Ṽ be an m-mixing stochastic isome-

try, ϕ1 , ϕ2 , ϕ3 ∈ H a triple of mutually orthogonal unit vectors, α, β ∈ C, |α| =
|β| = 1, k, ` ∈ {1, 2, 3}. Then:
1 + i (L) 1 − i (A)
5ϕ1 5ϕ1 +iϕ2 5ϕ1 +ϕ2 5ϕ1 = 5ϕ1 + 5ϕ1 ; (5.16)
4 4
1
5ϕ1 5ϕ1 +αϕ3 5ϕ1 +βϕ2 5ϕ1 = 5ϕ1 ; (5.17)
4
5ϕ1 5ϕ2 +αϕ3 5ϕk +βϕ` 5ϕ1 = 0; (5.18)
1 + i (L) 1 − i (A)
5ϕ1 5ϕ1 +iϕ3 5ϕ1 +ϕ3 5ϕ1 = eϕ +
5 eϕ .
5 (5.19)
1 1
4 4
Proof. We do not claim here that 5 e (L,A)

ϕ1 = 5(L,A)
ϕ1 , nor that these projections are
independent of the choice of ϕ2 , ϕ3 . This will be established in a later step.
Relations (5.16) and (5.19) are instances of Lemma 5.2. Due to 5ϕ2 +αϕ3 6
5ϕ2 + 5ϕ3 (cf. Equation (5.12)), we have 5ϕ1 5ϕ2 +αϕ3 = 0, which proves (5.18).
To verify (5.17), we compute:
5ϕ1 5ϕ1 +αϕ3 5ϕ1 +βϕ2 5ϕ1 = 5ϕ1 5ϕ1 +αϕ3 (5ϕ1 + 5ϕ2 + 5ϕ3 )5ϕ1 +βϕ2 5ϕ1
1
= 5ϕ1 5ϕ1 +αϕ3 5ϕ1 5ϕ1 +βϕ2 5ϕ1 = 5ϕ1 .
4
Here we have used the orthogonalities: 5ϕ1 +αϕ3 5ϕ2 = 5ϕ3 5ϕ1 +βϕ2 = 0. 2
96 P. BUSCH
Approaching the proof of Proposition 5.2, we observe that any triple of pairwise
independent unit vectors ϕ, ψ, ϑ spans a subspace [ϕ, ψ, ϑ] of H of dimension 2
or 3. We assume dim H > 3. As will be made evident, the case dim H = 2
can be dealt with by restriction of the constructions to be carried out for the case
dim H > 2. Let ϕ1 , ϕ2 , ϕ3 be three mutually orthogonal unit vectors such that
[ϕ, ψ, ϑ] ⊆ [ϕ1 , ϕ2 , ϕ3 ] =: K. Then the nine operators Aϕk ϕ` (cf. (5.7)) form a
basis of the space of operators on K. By polarization, an alternative basis is given
by the operators
Pϕ1 , Pϕ2 , Pϕ1 +ϕ2 , Pϕ1 +iϕ2 ,

Pϕ3 , Pϕ1 +ϕ3 , Pϕ2 +ϕ3 , Pϕ1 +iϕ3 , Pϕ2 +iϕ3 . (5.20)
We denote these Pξk , k = 1, . . . , 9, in the ordering given. It follows that the

corresponding set of projections 5ξk , k = 1, . . . , 9, i.e.
5ϕ1 , 5ϕ2 , 5ϕ1 +ϕ2 , 5ϕ1 +iϕ2 ,

5ϕ3 , 5ϕ1 +ϕ3 , 5ϕ2 +ϕ3 , 5ϕ1 +iϕ3 , 5ϕ2 +iϕ3 (5.21)
forms a basis of the space of operators spanned by all 5ϕ , ϕ ∈ K\{0}.
Proof of Proposition 5.2. Step 1. Excluding the trivial cases mentioned imme-
diately after Proposition 5.2, we need to verify Equation (5.9) for any triple of
unit vectors ϕ, ψ, ϑ ∈ H which are pairwise independent and nonorthogonal.
Consider the case dim H > 3. We work with a specific choice of ONB {ϕ1 , ϕ2 , ϕ3 }
of K, namely, ϕ1 = ϕ, ϕ2 ⊥ ϕ1 such that ψ ∈ [ϕ1 , ϕ2 ], and ϕ3 ⊥ [ϕ1 , ϕ2 ] such that
ϑ ∈ [ϕ1 , ϕ2 , ϕ3 ] = K. Using the operator bases (5.20) and (5.21), and denoting
these operators as Pξk , 5ξk in the order given, we can write
X
4 X
9
Pψ = sk Pξk , Pϑ = t` Pξ` ,
k=1 `=1
and
X
4 X
9
5ψ = sk 5ξk , 5ϑ = t` 5ξ` ,
k=1 `=1
with all sk , t` ∈ R. With this we find:
Pϕ Pϑ Pψ Pϕ = hϕ | ϑihϑ | ψihψ | ϕiPϕ

X
4 X
9
= sk t` Pϕ Pξ` Pξk Pϕ
k=1 `=1

1 1 1 1
= s1 t1 + s1 t3 + s1 t4 + s1 t6 + s1 t8 +
2 2 2 2
1 1 1+i 1 1
+ s3 t1 + s3 t3 + s3 t4 + s3 t6 + s3 t8 +
2 2 4 4 4
1 1−i 1 1 1
+ s4 t1 + s4 t3 + s4 t4 + s4 t6 + s4 t8 Pϕ . (5.22)
2 4 2 4 4
Next, applying Equations (5.16)–(5.18) (the expression (5.19) does not occur), we
evaluate the corresponding sum for the left-hand side of Equation (5.9):
X
4 X
9
5ϕ 5ϑ 5ψ 5ϕ = sk t` 5ϕ 5ξ` 5ξk 5ϕ = a5(L)
ϕ1 + ā5ϕ1 ,
(A)
(5.23)
k=1 `=1
where a turns out to be the same expression as that given in the braces {· · ·} in
Equation (5.22). Hence, a = hϕ | ϑihϑ | ψihψ | ϕi, and Equation (5.9) is verified.
Note that the case dim H = 2 is covered by putting sk = tk = 0 for k > 4.
Step 2. Next we show that the spectral projections occurring in Equation (5.9)
do not depend on the choice of ψ, ϑ ∈ H. Note that the result of Step 1 holds
for any choice of unit vectors ϑ ∈ H, and thus for any choice of ϕ3 ⊥ [ϕ1 , ϕ2 ] in
the case dim H > 3. The construction of 5(L) (A)
ϕ1 , 5ϕ1 so far depends on the choice
of ψ ∈ H, but only via ϕ2 ⊥ ϕ1 . In the case dim H = 2, the ray orthogonal to ϕ1
is uniquely determined by that vector, so that the spectral projections in Equation
(5.23) are the same for all ψ, ϑ. In the case dim H > 3, we have to show that these
projections are actually the same for all choices of unit vectors ϕ20 ⊥ ϕ1 (including
ϕ20 = ϕ3 , cf. Equation (5.19)). To this end we repeat the procedure of Step 1, this
time choosing a unit vector ϕ20 ⊥ ϕ1 = ϕ in the plane [ϕ, ϑ]. Then there exists
a unit vector ϕ30 such that ψ ∈ [ϕ1 , ϕ20 , ϕ30 ] = K. We obtain again a spectral
decomposition of the form (5.23),
0 0 0 0
5ϕ 5ϑ 5ψ 5ϕ = a5ϕ(L)
1
+ ā5ϕ(A)
1
, 5ϕ(L)
1
+ 5ϕ(A)
1
= 5ϕ , (5.24)
for the same set of vectors ϕ, ψ, ϑ. Provided that the eigenvalues a, ā do not coin-
0 0
cide, i.e. a ∈
/ R, then the spectral projections 5ϕ(L) 1
, 5ϕ(A)
1
constructed from ϕ1 , ϕ20
along the lines of the proof of Lemma 5.2 coincide with the projections 5(L) (A)
ϕ1 , 5ϕ1
constructed from ϕ1 , ϕ2 .
We show that, given ϕ, ψ, any ϕ20 ⊥ ϕ1 can be realized with a choice of ϑ such
that a = hϕ1 | ϑihϑ | ψihψ | ϕ1 i ∈ / R. Let ϕ20 be any unit vector such that ϕ20 ⊥ ϕ
0
and ϕ2 ⊥ / ψ. Choose ϑ = α1 ϕ1 + α2 ϕ20 , with α1 , α2 to be further specified shortly.
Take a unit vector ϕ30 orthogonal to ϕ1 , ϕ20 and such that ψ = β1 ϕ1 + β2 ϕ20 + β3 ϕ30 .
Since ϕ20 ⊥/ ψ, we have β1 6= 0 6= β2 . We obtain
a = α1 {ᾱ1 β1 + ᾱ2 β2 }β̄1 = |α1 |2 |β1 |2 + α1 ᾱ2 β2 β1 .
The first term on the right-hand side is real and positive. Then a 6= R can be easily
achieved by proper choice of α1 , α2 . Hence all choices of ϕ20 ⊥ ϕ1 , ϕ20 6= ϕ3 lead to
the same spectral decomposition in Equations (5.10) or (5.16). (Note that ϕ20 = ϕ3
would imply β2 = 0.)
98 P. BUSCH
A continuity argument finally shows that the case ϕ20 = ϕ3 can be included,
too, so that the spectral decompositions in Equations (5.16) and (5.19) do indeed
coincide. In fact, let ϕ2(n) be a sequence of unit vectors orthogonal to ϕ1 and different
from ϕ3 , such that kϕ2(n) − ϕ3 k → 0. Then Pϕ1 +αϕ(n) converges to Pϕ1 +αϕ3 in trace
2
norm. By the continuity of T , 5ϕ1 +αϕ(n) converges to 5ϕ1 +αϕ3 in trace norm. Then
2
1 + i (L) 1 − i (A)
5ϕ1 + 5ϕ1
4 4
= 5ϕ1 5ϕ1 +iϕ(n) 5ϕ1 +ϕ(n) 5ϕ1 −→ 5ϕ1 5ϕ1 +iϕ3 5ϕ1 +ϕ3 5ϕ1 ,
2 2
convergence in trace norm. It follows that the first and last expressions coincide.
This concludes Step 2.
We introduce the following notation for the operators of Equations (5.9):
ϕ ϕ ϕ
Wϑψ = 5ϕ 5ϑ 5ψ 5ϕ = aϑψ 5(L)
ϕ + aϑψ 5ϕ ,
(A)
ϕ
aϑψ = hϕ | ϑihϑ | ψihψ | ϕi.
PROPOSITION 5.3. Let T : V → Ṽ be an m-mixing stochastic isometry. For any

two unit vectors ϕ, χ ∈ H one has
ϕ 5χ = hPχ , Pϕ i5χ , 5χ 5ϕ 5χ = hPχ , Pϕ i5χ ;

5χ 5(L) (L) (A) (A)
(5.25)
5(L)
ϕ 5χ
(A)
= 5(A)
ϕ 5χ
(L)
= 0. (5.26)
The projections
_ _
P (L) = 5(L)
ϕ , P (A) = 5(A)
ϕ (5.27)
ϕ∈H\{0} ϕ∈H\{0}
are mutually orthogonal. Moreover, for all ϕ ∈ H\{0},
ϕ =P
5(L) ϕ =P
(L)
5ϕ , 5(A) (A)
5ϕ , (5.28)
and the ranks m(L) , m(A) of the projections 5(L) (A)

ϕ , 5ϕ are independent of ϕ ∈
H\{0} and satisfy m + m = m. If the invariants m(L) , m(A) are nonzero, then
(L) (A)
T decomposes into a convex combination of two stochastic isometries T (L) , T (A)

with orthogonal ranges,
m(L) (L) m(A) (A)
T = T + T , (5.29)
m m
1 1
T (L) (ρ) = P (L) T (ρ)P (L) , T (A) (ρ) = P (A) T (ρ)P (A) . (5.30)
m(L) m(A)
Proof. Let χ ∈ H be a unit vector not orthogonal to ϕ, and choose any unit
ϕ
vector ψ ∈ H neither orthogonal nor parallel to ϕ, χ and such that aχψ ∈
/ R.
ϕ
We compute the operator 5χ Wχψ 5χ in two ways, using (5.3) and noting that
ϕ χ
aχψ = aψϕ :
ϕ χ ϕ ϕ
5χ Wχψ 5χ = hPχ , Pϕ iWψϕ = hPχ , Pϕ i aχψ 5(L)
χ + aχψ 5χ
(A)
ϕ ϕ
= aχψ ϕ 5χ + aχψ 5χ 5ϕ 5χ .
5χ 5(L) (A)
Using the fact that the map 5χ |5ϕ H : 5ϕ H → 5χ H is orthogonality preserving

(cf. Equation (5.4)), Equations (5.25) and (5.26) follow by application of Equation
(5.2) and a version of (5.4). The orthogonality of P (L) , P (A) and the relation (5.28)
are then obvious.
The invariance of the numbers m(L), m(A) is a consequence of the fact that
5χ transforms an orthogonal basis of 5(L) (A)
ϕ H[5ϕ H] onto an orthogonal basis
of 5(L) (A)
χ H[5χ H].
For ϕ ∈ H\{0} we have, by virtue of Equation (5.28):
1 1 (L)
T (Pϕ ) = 5ϕ = P 5ϕ P (L) + P (A) 5ϕ P (A)
m m
(L)
m m(A) (A)
= T (L) (Pϕ ) + T (Pϕ ).
m m
Due to the linearity and continuity of T this equation extends to all (σ -)convex
combinations of pure states and thus to all states ρ and finally to all of V . 2
Next we come to analyze the stochastic isometries T (L) , T (A) , or equivalently,

the m-mixing stochastic isometries for which either m = m(L) or m = m(A) . We
note that
ϕ ϕ
Wϑψ = 5ϕ 5ϑ 5ψ 5ϕ = aϑψ 5ϕ [case m = m(L)], (5.31)
ϕ ϕ
Wϑψ = 5ϕ 5ϑ 5ψ 5ϕ = aϑψ 5ϕ [case m = m(A) ]. (5.32)
Let T be an m-mixing stochastic isometry with m = m(L) or m = m(A) . Pick

an arbitrary vector ϕ ∈ H\{0} and an orthogonal basis {ϕ1 , . . . , ϕm } of 5ϕ H. For
any ψ ∈ H\{0} with ψ ⊥ / ϕ, the set of vectors
ψk := hPψ , Pϕ i−1/2 5ψ ϕk (5.33)
is an orthonormal basis of 5ψ H, cf. Equation (5.4). Moreover, for any two unit
vectors ψ, ϑ not orthogonal to ϕ, the corresponding basis vectors ψk , ϑ` are mu-
tually orthogonal if k 6= `. In fact, due to Equations (5.31), (5.32) we have (∼
denoting proportionality)

ϕ
hϑ` | ψk i ∼ h5ϑ ϕ` | 5ψ ϕk i = ϕk | Wϑψ ϕ` ∼ δk` . (5.34)
100 P. BUSCH
PROPOSITION 5.4. Let T be an m-mixing stochastic isometry with m = m(L) or

m = m(A) . Fix a vector ϕ 0 ∈ H\{0} and an orthonormal basis {ϕ1 , . . . , ϕm } of
5ϕ0 H. The projections
_
Pk := Pψk (5.35)
ψ⊥
/ ϕ0
are mutually orthogonal. For any ψ ∈ H\{0} there exists a unique ray Pψk H, ψk ∈
5ψ H, such that
Pk 5ψ = 5ψ Pk = Pψk , k = 1, . . . , m. (5.36)
Then
_
Pk = Pψk , k = 1, . . . , m, (5.37)
ψ∈H\{0}
and
X
m _
Pk = 5ϕ . (5.38)
k=1 ϕ∈H\{0}
T decomposes into a convex combination of pure stochastic isometries,

Xm
1
T = Tk , (5.39)
k=1
m
Tk (ρ) = mPk T (ρ)Pk . (5.40)
Hence there exist maps Uk : H → Pk H̃, k = 1, . . . , m, which are all unitary if

m = m(L) or all antiunitary if m = m(A) , such that Pk = Uk Uk∗ and
Tk (ρ) = Uk ρUk∗ , ρ ∈ B1 (H)s . (5.41)
Proof. The orthogonality of the Pk is a direct consequence of Equation (5.34).

Let ψ be a nonzero vector. If ψ ⊥ / ϕ 0 , then take ψk ∼ 5ψ ϕk . We have Pψk 6 Pk
for all k, thus Pψk Pk = Pψk and,Pdue to the orthogonality of the Pk , Pψ` Pk = 0 if
` 6= k. We obtain Pk 5ψ = Pk ` Pψ` = Pk Pψk = Pψk . This proves (5.36) for the
case ψ ⊥/ ϕ 0.
Consider the case ψ ⊥ ϕ 0 . We assume that ϕ 0 , ψ are unit vectors. Define a
sequence of unit vectors ϑ (n) := ( n1 )1/2 ϕ 0 + (1 − n1 )1/2 ψ → ψ as n → ∞. We
can define an ONB of 5ϑ (n) H as in (5.33), ϑk(n) ∼ 5ϑ (n) ϕk . Using again Equations
(5.31), (5.32), it is easy to verify that for each k, the vectors ϑk(n) form a Cauchy
sequence. Let ψk denote the limiting unit vector. We show that Pk 5ψ = Pψk . (This
ensures the uniqueness of the ray Pψk .)
Let ξ ∈ H be any vector. One estimates

P (n) − Pψ ξ 6 kϑ (n) − ψk k |hϑ (n) | ξ i| + kψk k |hϑ (n) − ψk | ξ i|
ϑ k
k k k k
6 2kϑk(n) − ψk k kξ k.
We conclude that kPϑ (n) − Pψk k → 0 and so kPϑ (n) − Pψk k1 → 0. We also have
k k
kPϑ (n) − Pψ k1 → 0 and therefore, due to the continuity of T , k5ϑ (n) − 5ψ k1 → 0
and also kPk 5ϑ (n) − Pk 5ψ k1 → 0. But Pk 5ϑ (n) = Pϑ (n) → Pψk , and so Pk 5ψ =
k
Pψk . This proves (5.36).
The relation (5.37) is an immediate consequence
P of the fact just demonstrated
P
that for all ψ, Pψk 6 Pk . Finally we have ( k Pk )5ϕ = 5ϕ and so 5ϕ 6 k Pk .
Since all Pψk 6 5ψ , the converse ordering holds as well. This proves (5.38).
The maps Tk are clearly linear and positive. We have Tk (Pξ ) = Pk 5ξ = Pξk for
all vectors ξ 6= 0, so the Tk are trace preserving and thus pure stochastic maps. Let
ψ, ξ ∈ H\{0}, with ψ ⊥ ξ . Then 5ψ ⊥5ξ (since T is orthogonality preserving),
and so Pψk ⊥Pξk , that is, Tk (Pψ )⊥Tk (Pξ ). Thus the Tk are orthogonality preserv-
ing and therefore (pure stochastic) isometries. Equation (5.39) then follows by
straightforward application of (5.40). The existence of unitary or antiunitary maps
Uk satisfying (5.41) is a consequence of Theorem 2.3.1 of [5] (cf. Proposition 3.1
above).
If Uk is unitary, we find, using Uk∗ Uk = I :
Tk (Pϕ )Tk (Pϑ )Tk (Pψ )Tk (Pϕ ) = Uk (Pϕ Pϑ Pψ Pϕ )Uk∗ = aϑψ Tk (Pϕ ).
ϕ
Similarly, if Uk is antiunitary, we find:
Tk (Pϕ )Tk (Pϑ )Tk (Pψ )Tk (Pϕ ) = Uk (Pϕ Pϑ Pψ Pϕ )Uk∗ = aϑψ ϕ
Tk (Pϕ ).
P
Observing that Tk (Pξ )⊥T` (Pχ )(k 6= `) and 5ϕ = Tk (Pϕ ), etc., we obtain:
ϕ
X X
Wϑψ = 5ϕ 5ϑ 5ψ 5ϕ = Tk (Pϕ )Tk (Pϑ )Tk (Pψ )Tk (Pϕ ) = ak Tk (Pϕ ),
k k
ϕ ϕ
where ak = aϑψ or ak = aϑψ according to whether Uk is unitary or antiunitary.
Comparing this with Equations (5.31), (5.32), it is seen that all the Uk must either
be simultaneously unitary or antiunitary in order to reproduce the right-hand sides
of these equations.
From Tk (Pϕ ) · T` (Pψ ) = 0 for k 6= `, ϕ, ψ ∈ H\{0}, it follows that Uk∗ U` =
δk` IH . Also, it is obvious that the projection Uk Uk∗ has the same range as Pk , and
so Uk Uk∗ = Pk . 2
REMARK 5.1. The stochastic isometries T (L), T (A) of Proposition 5.3, Equa-
tion (5.30), are convex combinations of pure stochastic isometries generated by
linear and antilinear isometries of H, respectively. This explains the use of the
superscripts (L), (A) throughout this section.
102 P. BUSCH
Proofs of Proposition 3.4 and Theorem 1. Proposition 3.4 is a direct conse-

quence of the combination of Propositions 5.3 and 5.4. Theorem 1 is a direct result
of the combination of Propositions 3.3 and 3.4. 2
P
Proof of Theorem 2. Let T = k wk Tk be a stochastic isometry of the form
of Equation (3.2), the Tk being pure stochastic isometries generated by unitary or
antiunitary maps Uk : H → H̃k . For Uk unitary (antiunitary), Tk is completely
positive (not completely positive). The sum of completely positive maps is com-
pletely positive, and this extends to σ -convex combinations. Hence T is completely
positive whenever all Uk are unitary.
Conversely, let T be completely positive. Suppose some Uk is antiunitary. Hence
there exist unit vectors 9 ∈ H̃k ⊗ Cn , 2 ∈ H ⊗ Cn for some n ∈ N such that
h9 | T ⊗ ι(P2 )9i = h9 | Tk ⊗ ι(P2 )9i < 0. (5.42)
(Here ι denotes the identity map on B1 (Cn ).) The equality is due to the choice of
9 in the subspace H̃k ⊗ Cn and the fact that the ranges of the U` are mutually
orthogonal. Equation (5.42) contradicts the complete positivity of T . 2
6. Concluding Remarks. Some Physical Applications

Stochastic isometries give rise to a variety of associated maps some of which will
be briefly described here. Any linear bounded map T : B1 (H)s → B1 (H̃)s has a
unique extension to a linear bounded map on B1 (H). This extension of T will be
denoted T̂ . A stochastic map T has a unique extension to a linear, trace preserving
map on B1 (H).
The extension T̂ to B1 (H) of a pure stochastic isometry T generated by a
unitary or antiunitary map U : H → H̃ is given as follows: for ρ ∈ B1 (H),
T̂ (ρ) = UρU ∗ if U is unitary,

T̂ (ρ) = Uρ ∗ U ∗ if U is antiunitary.
PROPOSITION 6.1. The extension T̂ : B1 (H) → B1 (H̃) of a stochastic isom-

etry T : B1 (H)s → B1 (H̃)s is a trace preserving isometry. Conversely, the
restriction to B1 (H)s of a trace preserving isometry on B1 (H) is a stochastic
isometry.
Proof. Let T be a stochastic isometry, T̂ its extension. That T̂ is trace preserving
follows triviallyP
from the corresponding property or T . Using Equation (3.2), one
finds |T (ρ)| = k wk |Uk ρUk∗ | for ρ ∈ B1 (H), and so
X
T̂ (ρ) = tr |T (ρ)| = wk kρk1 = kρk1 .
1
k
Conversely, assume T̂ is a trace preserving isometry on B1 (H), and let T denote

its restriction to B1 (H)s . T is clearly trace preserving and isometric and therefore,
by Lemma 2.1, a stochastic isometry.
P
PROPOSITION 6.2. Let T = k wk Tk : V → Ṽ be a stochastic isometry, with
Tk pure stochastic isometries as in Equation (3.2). Then
X√
T̃ (τ ) := wk Tk (τ ), τ ∈ B2 (H),
k
defines an isometry T̃ of the Hilbert–Schmidt class B2 (H).

Proof. For τ ∈ B2 (H), one easily verifies that
X
T̃ (τ )∗ T̃ (τ ) = wk Tk (τ )∗ Tk (τ ) ∈ B1 (H)
k
(where we used the orthogonality Tk (τ )∗ T` (τ ) = 0, k 6= `), and so

X
T̃ (τ ) 2 = wk tr [Uk τ ∗ τ Uk∗ ] = τ k22 .
2
k
Therefore, kT̃ (τ )k2 = kτ k2 . 2

P
REMARK 6.1. Let T = k wk Tk be a stochastic isometry, with pure stochastic
P
isometric components Tk , T̂ = k wk T̂k its extension to B1 (H). It follows that
◦
P
T := k T̂k is an isometry on B(H) and a Jordan ∗-isomorphism, and that
its restriction to the orthocomplemented lattice of projections of H, P (H), is a
lattice- and ortho-isomorphism from P (H) onto the lattice
PT := {5(T (ρ)) : ρ ∈ B1 (H)s }.
The map ◦ T decomposes in a unique way into

◦
T = ◦ T (L) + ◦ T (A) ,
◦ (L)
T (a) = P (L)◦ T (a)P (L) , ◦ T (A) (a) = P (A)◦ T (a)P (A) , a ∈ B(H),
and ◦ T (L) (◦ T (A) ) is a ∗-isomorphism (∗-anti-isomorphism) of the C ∗ -algebra B(H)
onto the subalgebras ◦ T (L) (B(H))(◦ T (A) (B(H))). This is an illustration of the
result of Kadison [8] cited in Section 1.
Further we note that the projections P (L) , P (A) are in the centre of the von
Neumann algebra generated by ◦ T (B(H)). In physical terms, they induce a super-
selection rule on the state space T (B1 (H)s ).
REMARK P 6.2. Consider a completely positive m-mixing stochastic isometry,

∗
T = m1 m T
k=1 k , with Tk (ρ) = U k ρU k , all Uk unitary (hence m(L) = m). The maps
Uk are not uniquely determined by T , even apart from a phase factor, if m > 2.
104 P. BUSCH
This is apparent from the construction of the pure isometries Tk in Proposition 5.4,
which was based on the choice of some nonzero vector ϕ 0 ∈ H and Pan arbitrary
ONB ϕk of 5ϕ0 H. Accordingly, it can be shown that if T (ρ) = m1 k Uk ρUk∗ =
1
P ∗
m k Vk ρVk , where Vk is another set of unitary maps, then
X
Vk = γk` U` ,
`
with (γk` ) a unitary m × m matrix. It is easily verified that this condition ensures
that the Vk define the same mixing stochastic isometry as the Uk . The totality of
all projections Qk = Vk Vk∗ thus obtained from a given set Uk using all unitary
matrices (γk` ) commute with all elements of T (B1 (H)) and ◦ T (B(H)).
REMARK 6.3. The relation 5ϕ 5ψ 5ϕ = α5ϕ , valid for m-mixing stochastic

isometries shows that the set of projections 5ϕ bears some fundamental similarities
with the set of one-dimensional projections. In fact pairs of projections satisfying
such a relation (with nonzero factor α) are called isoclinic; their geometric signifi-
cance has been studied by von Neumann [9]. Note that also the projections Pk , Q`
discussed in Remark 6.2 are isoclinic.
A physically interesting feature of a stochastic isometry T lies in the fact that

its inverse on the range T (V ) can be extended to a stochastic map on V .
THEOREM 3. Let T : V → Ṽ be a stochastic isometry. The inverse T −1 :

T (V ) → V admits an extension to a stochastic map on Ṽ as follows. For T
expressed in the form of Equation (3.2), let P0 be the projection onto H̃0 . The
following defines a stochastic map on Ṽ :
X
N
S(ρ) = Uk∗ ρUk + P0 ρP0 , ρ ∈ Ṽ .
k=1
Then T −1 = S|T (V ) .
Proof. The positivity of S is obvious. TracePpreservation follows from the fact
that Uk Uk∗ = Pk is the projection onto H̃k and Nk=0 Pk = IH̃ .
∗
Let ρ = T (σ ). Then, since Uk U` = δk` IH and P0 ρP0 = 0, one has
X
N X
N X
N
S(ρ) = S(T (σ )) = wk U`∗ Uk σ Uk∗ U` = wk σ = σ = T −1 (ρ).
`=1 k=1 k=1
Hence S|T (V ) = T −1 . 2
This result suggests the possibility that reversible physical state changes are not
necessarily represented by surjective, and hence pure, stochastic isometries, but
that a state change effected by any stochastic isometry is reversible: there exists
a single dynamical map that sends all final states back to the initial states. This
interpretation is further elaborated in [3].
Pursuing further the dynamical interpretation of stochastic isometries, it may
be observed that such maps induce a reduction of the symmetries of the physical
system in question. In this sense it can be said that dynamical maps represented
as stochastic isometries describe the formation of structure (in a reversible way).
According to a theorem due to Wigner, any symmetry operation [1], defined as an
angle preserving map of the set of rays of H onto itself, is induced by a unitary
or antiunitary map according to Equation (3.1). For simplicity, we consider a com-
pletely positive stochastic isometry of the form (3.2), with all Uk unitary. Let U be
a unitary symmetry operation. Then we compute:
X
N
T (UρU ∗ ) = wk (Uk U Uk∗ )Uk ρUk∗ (Uk U ∗ Uk∗ ) = Ũ T (ρ)Ũ ∗ ,
k=1
where
X
Ũ = Uk U Uk∗ = ◦ T (U )
k
is unitary. The map U 7→ Ũ is injective. But to every U there do exist unitary maps
V 6= Ũ such that
Ũ T (ρ)Ũ ∗ = V T (ρ)V ∗ (6.1)
PN
for all ρ ∈ B1 (H). In fact, define a unitary map W = k=0 λk Pk , with Pk =
∗
Uk Uk , λk ∈ C, |λk | = 1, such that W 6= I . Let V = Ũ W . Then (6.1) holds
due to the fact that W commutes with all T (ρ). This shows that on the state space
T (B1 (H)s ) not all symmetries of H can be distinguished.
There is another interpretation of a stochastic isometry T and its associated map
◦
T defined in Remark 6.1. These maps lead to a physically equivalent description
of all states and observables of the given quantum system in the following sense:
for all states ρ and effects a, the corresponding states T (ρ) and effects ◦ T (a) give
the same probabilities:
hT (ρ), ◦ T (a)i = hρ, ai.
This observation can be elaborated into a general theory of extensions of the statis-
tical description of a quantum physical system on the basis of Theorem 1 [2].
Appendix A. Algebraic Proof of Proposition 3.4

The decomposition of a mixing stochastic isometry into pure stochastic isometries
can be obtained by application of Kadison’s work on isometries of operator alge-
bras [8], thereby bypassing the explicit geometric constructions of Section 5. Here
we sketch the required steps.
106 P. BUSCH
Let T : V → Ṽ be an m-mixing stochastic isometry. One first shows that the

map ◦ T : V → Ṽ , ◦ T = mT extends to a linear, ∗-preserving, norm-bounded
map ◦ T̂ : B(H) → B(H̃). This map sends projections to projections and is
σ -ortho-additive on the set of projections.
Following an argument of Wright [10], ◦ T̂ is finally shown to be a Jordan
∗-homomorphism.
The decomposition of ◦ T̂ , and hence of T , is then obtained by application of
the arguments of Wright [10], which make use of Kadison’s theorem on isometries
[8] in the form presented by Emch ([6], Theorem 1, p. 153).
Acknowledgement
The author wishes to thank Pekka Lahti for helpful comments on an earlier version
of the manuscript.
References
1. Bargmann, V.: Notes on Wigner’s theorem on symmetry operations, J. Math. Phys. 5 (1964),
862–868.
2. Busch, P.: Quantum extensions of quantum statistical models, Preprint, 1998.
3. Busch, P. and Quadt, R.: Operational characterization of irreversibility, Report Series, Depart-
ment of Mathematics, University of Hull, 1998.
4. Cassinelli, G., DeVito, E., Lahti, P. J. and Levrero, A.: Symmetry groups in quantum mechanics
and the theorem of Wigner on the symmetry transformations, Rev. Math. Phys. 9 (1997), 921–
941.
5. Davies, E. B.: Quantum Theory of Open Systems, Academic Press, New York, 1976.
6. Emch, G. G.: Algebraic Methods in Statistical Mechanics and Quantum Field Theory, Wiley,
New York, 1972.
7. Kadison, R. V.: Transformations of states in operator theory and dynamics, Topology 3 (Suppl.
2) (1965), 177–198.
8. Kadison, R. V.: Isometries of operator algebras, Ann. of Math. 54 (1951), 325–338.
9. Maeda, S.: Probability measures on projections in von Neumann algebras, Rev. Math. Phys. 1
(1990), 235–290.
10. Wright, R.: The structure of projection-valued states: a generalization of Wigner’s theorem, Int.
J. Theor. Phys. 16 (1977), 567–573.
113
Complex Star Algebras
L. BOUTET DE MONVEL
Institut de Mathématiques de Jussieu, Université Pierre et Marie Curie, 4 place Jussieu,
75251 Paris Cedex 05, France; e-mail: boutet@math.jussieu.fr
(Received : 17 November 1998)

Abstract. We describe a classification of star algebras on the cotangent bundle of a complex man-
ifold, locally isomorphic to the algebra of pseudo-differential operators; this requires a slight exten-
sion of the usual definition of star algebras. We show that in dimension > 3 these are essentially
trivial and come from algebras of differential operators on X; in dimension 1 and 2 there are many
more, which we describe.
Mathematics Subject Classifications (1991): 16S80, 16S32, 57D17.
Key words: star products, deformation, holomorphic.
1. Introduction
Let us first recall what a star product is (detailed definitions are given inPSection 2).
Let X be a manifold and let O b denote the algebra of formal series f = k>k fk hk ,
0
where the fk are smooth functions on X and h is a “small” formal parameter. A star
product on X is a unitary algebra law on O b for which the unit is 1 and the product
is local, i.e., given by a formula
X
f, g → B(f, g) = fg + hk Bk (f, g),
k >k0
where the Bk are bidifferential operators on X: in local coordinates Bk (f, g) =

P
aαβ ∂ α f ∂ β g with smooth coefficients aαβ (it is further required that the unit is
1, i.e., B0 (f, g) = fg and Bk (1, f ) = Bk (f, 1) = 0 for any k > 0 and any
f ; the addition law is the usual addition of O. b A star product can be thought
of as a noncommutative deformation of the usual product. The leading term of
commutators {f, g} = hB1 (f, g) − hB1 (g, f ) defines a Poisson bracket on X (star
products are also called “deformation quantization of Poisson manifolds”).
In this paper, I will use a slightly extended definition, where star products live on
cones. A cone 6 with basis B6 = X is the complement of the zero section in a line
bundle L → X (a complex line bundle if X is a complex manifold, and preferably
a half-line bundle if X is real); in the semi-classical case above 6 = X × R+ and
h = 1/r if r denotes the fiber variable (the small “Planck constant” plays the role
of the inverse of a large frequency). In this context O b is the set of formal series
114 L. BOUTET DE MONVEL
P
f = k6k0 fk where for each k, fk is a function homogeneous of degree k on 6
b
and, locally, a star product is defined as above as a bidifferential product law on O
X
f, g → B(f, g) = fg + Bk (f, g),
k 6k0
where Bk is now a bidifferential operator on 6, homogeneous of degree k → −∞

with respect to fiber homotheties. The Bk may involve derivations in any direction,
so there is no longer a distinguished “Planck constant” commuting with the rest? .
The associated Poisson bracket now lives on 6 and is homogeneous of degree −1.
This definition includes the algebras of pseudo-differential operators or Toeplitz
operators, which are after all among the most important and belong to the same
formalism.
Complex star algebras arrive naturally and unavoidably in many problems con-
cerning differential operators, whose symbols are polynomials and always live on
a complex manifold. So it is important to study them, and to study their relations
with “polynomial” objects associated to differential operators.
In his paper [22] M. Kontsevitch has shown that any homogeneous Poisson
bracket on a real manifold comes from a star product. His proofs extend without
changing a word to star products on a cone. Kontsevitch’s formula giving a star
product from a Poisson bracket on an affine space also works without any modi-
fication in the complex case (i.e., 6 = Cn × C× ). But the argument used to go
from local to global does not work for complex manifolds, because it uses in an
unavoidable manner partitions of unity and tubular neighborhoods. In general, I do
not know if a global star product exists for a given Poisson bracket, even in the
symplectic case, nor do I know what the classification of such algebras looks like
(see, however, [20] where it is shown that even if such an algebra E may not exist,
the category of sheaves of E-modules can be defined up to equivalence).
In this paper, I investigate those star algebras which live on a complex cotangent
cone T ∗ X−{0} deprived of its zero section, equipped with its canonical symplectic
Poisson bracket. All star algebras associated with this Poisson bracket are locally
isomorphic, and there exists a global such algebra, viz. the algebra of pseudo-
differential operators; so there is at least a starting point for the classification.
This will turn out to be essentially trivial in dimension n > 3 (Theorem 1),
but instructively not in dimension 2. Algebras over a manifold X of dimension
2 or > 2 are described in Section 4 more precisely and compared to D-algebras,
i.e., sheaves of algebras over X locally isomorphic to E, the algebra of pseudo-
differential operators coming from differential operators on X. It turns out that
if dim X > 3 we get nothing new: the functor which takes a D-algebra to the
associated star-algebra is an equivalence. If dim X = 2 the same functor is faithful,
i.e., two D-algebras are isomorphic if and only if the associated star algebras are
isomorphic, and an isomorphism between such star algebras comes from a unique
? There is absolutely no reason that the Planck constant should commute with the rest, especially
when it is a parameter without physical significance.
COMPLEX STAR ALGEBRAS 115
isomorphism between the original D-algebras; however, there are in general many
more “exotic” star algebras which do not come from a D-algebra.
If X is of dimension 1 the classification depends on whether X is open, of genus
> 2, of genus 1 or of genus 0.
An inner automorphism of the algebra E of pseudo-differential operators on
X(U : P → AP A−1 ) has a symbol σ (U ) = d Log σ (A), which is a section of
the sheaf ω (on the “basis” B6 = 6/C× of closed forms homogeneous of degree
0 on 6, and an exponent which is the degree of A; we will see in Section 2 that
any automorphism U of E has likewise a symbol and an exponent ∈ C. Similarly,
a star algebras has a symbol σ (A) ∈ H 1 (B6, ω) and an exponent ∈ H 1 (B6, C).
We will see in Section 3 that if X is an open curve or a curve of genus > 1, star
algebras on 6 are completely determined by their exponent. The classification is
more subtle when X is closed of genus 1 or 0.
The techniques used in this paper are a mixture of noncommutative cohomol-
ogy, holomorphic cohomology, and the relation between the cohomology of a sheaf
with a filtration and the cohomology of the associated graded sheaf. This contains
nothing really new or difficult, but the mixture can be somewhat muddling.
As far as I know the questions studied here have not been investigated before
and the results are new.
In Sections 2 and 3 we recall the definition of star algebras, and some classifi-
cation principles.
In Section 4 we describe the classification when dim X > 2.
In Section 5 we describe the case where X is a curve (dim X = 1): results are
substantially different if X is open, X = P1 , X is of genus 1, or X is of genus
g > 2.
2. Star Algebras
2.1. CONES
DEFINITION 1. A real (resp. complex) cone is a C ∞ (resp. holomorphic) prin-

cipal bundle 6 with group R× × ×
+ (resp. C ). The basis is B6 = 6/R+ (resp.
6/C×).
A real cone is isomorphic to a product cone B6 × R× ?

+ . A complex cone is
isomorphic to L − {0} (L deprived of its zero section) where L is a complex line
bundle over B6. L will usually not be a trivial bundle.
DEFINITION 2. (i) We denote O(m) the sheaf on B6 of homogeneous functions

of degree m of 6 (holomorphic in the complex case).
? At least if we are dealing with paracompact manifolds, which will always be the case in this
article.
b the sheaf on B6 of formal symbols (“asymptotic expansions”

(ii) We denote O
for ξ → ∞ in 6):
X
f ∈O b if f = fm with fm ∈ O(m) (1)
m6m0
(m an integer, m → −∞).
DEFINITION 3. For an integer k > 1 we denote b

P Dk the sheaf (on B6) of formal
k-differential operators: P (f1 , . . . , fk ) = m6m0 Pm (f1 , . . . , fk ) with Pm a k-
linear differential operator homogeneous of degree m with respect to homotheties
(m an integer, m → −∞).
b
If k = 1 we will just write D.
Locally 6 is a product cone and we may choose homogeneous coordinates (real

or complex) xj of degree 0 on the basis, and r of degree 1 on the fiber. Then
Pm (f1 , . . . , fk ) is a sum of monomials
ϕ(x)r m ∂xα1 (r∂r )m1 (f1 ) . . . ∂xαk (r∂r )mk (fk ).
There is no restriction on the order of Pm .

The presence of two “degrees” is confusing so in what follows degree will
always refer to the degree with respect to homotheties, and order refers to the
degree as a differential operator; thus if P ∈ D bk each term Pm of degree m is of
finite order, although the resulting infinite sum P may be of infinite order.
WePwill denote D b× ⊂ D b the sheaf of invertible formal differential operators:
P = Pk ∈ D b is invertible iff its leading term σ (P ) = Pm0 is invertible, i.e.,
×
Pm0 is of order 0, the multiplication by a nonvanishing function homogeneous of

degree m0 . We denote by D b−× the subsheaf of those invertible P such that P (1) =
1, i.e., P is of degree 0, its leading term is P0 = 1 and terms of lower degree have
no constant term: Pm (1) = 0 if m < 0.
Remark 1. Sheaves are of course useless in the real case but must be used in
the complex case where global sections do not necessarily exist.
Remark 2. For analytic cones there is also a notion of convergent symbol (in-
troduced by the author in [6] to define analytic pseudo-differential operators).
These are in fact the more important and for many questions it is essential to
use convergent rather than formal symbols.? However, for the classification results
below, there is no significant qualitative difference between formal and convergent
symbols, so we will stick to formal symbols and avoid convergence technicalities.
? e.g., convergent rather than formal symbols are essential in the finiteness theorems of T. Kawai
and M. Kashiwara [22], or for going from E -modules to D-modules in the thesis of D. Meyer [23],
and probably in most problems involving a comparison between E and D-modules.
2.2. STAR PRODUCTS ON A REAL OR COMPLEX CONE
DEFINITION 4. A star product on 6 is a sheaf A on the basis B6, locally

isomorphic to O b as a sheaf of vector spaces (the structural sheaf of groups is
described below), equipped with an associative unitary algebra law whose product
(star product) f ∗ g = B(f, g) is locally a formal bidifferential operator.
Locally f ∗ g = 6Bm (f, g) with Bm a bidifferential operator homogeneous of
degree m → −∞, B0 = 1. The first idea is that the structural sheaf of groups used
to patch together local frames of A is the sheaf D b× (on B6) of invertible formal
differential operators, but there is a unit that we can choose equal to 1 in all local
frames so this obviously reduces to D b−× .
Note that homotheties (hence degrees) are not respected by D b−× . However, if
P ∈D b− , f and Pf have the same leading term; so P respects the filtration defined
×
P L
by degrees (f ∈ O bm if f = j 6m fj ) and gr P is the identity on gr O b= O(m).
In the semi-classical definition, 6 is a product cone 6 = B6 × L (L = R+ or
C× ), the star product law is defined on O b and does not involve vertical derivatives,
−1
so the “Planck constant” h = r plays the role of a constant. The definition above
includes the “semi-classical” case and also the algebras of pseudo-differential or
Toeplitz operators. This conic framework for star products was described in [4].
In the real case, using partitions of unity, it is immediate to see that A is always
isomorphic to O b as a sheaf (“there exists a global total symbolic calculus”). This is
no longer true in the complex case, and in particular it is not true in the most simple
and natural examples as we will see below, so the sheaf theoretic presentation
cannot be avoided.
2.3. ASSOCIATED POISSON BRACKET
If A is a star algebra on 6 it has a canonical filtration coming from the filtration

b by homogeneity degrees, and there is a canonical isomorphism gr A ' O
of O b
because the structural sheaf of groups Db− induces the identity on gr O.
× b The com-
mutator law then defines a Poisson structure on gr A = O, b i.e., the leading term of
the commutator law
{f, g} = B1 (f, g) − B1 (g, f )
is a Poisson bracket on 6, homogeneous of degree −1. This means that it is a
skew-symmetric bivector field
{f, g} = −{g, f }, {f, gh} = {f, g}h + g{f, h}
satisfying the Jacobi identity (i.e., it is a Lie bracket):
{f {g, h}} = {{f, g}h} + {g{f, h}}
and it is homogeneous of degree −1 with respect to homotheties.
deg {f, g} = deg f + deg g − 1 if f, g are homogeneous.
Existence of a global star algebra on a real symplectic cone 6 was proved by

V . Guillemin and myself in [5] (see also [3]), and by M. De Wilde and P. Lecomte
([9, 10]) in the semiclassical symplectic case (cf. also the nice deformation proof
of B. V. Fedosov [11]).
In [22] M. Kontsevitch proved that any Poisson bracket comes from a star prod-
uct in the real semiclassical case. More precisely, he proves that there is a one to
one correspondence between isomorphic classes of star products and isomorphic
classes of formal families of Poisson brackets depending on the “small parameter”
h. His result extends, without changing a word, to star products on a real cone with
the definition above; families of Poisson brackets should be replaced by formal
Poisson brackets on 6:
X
c= cm . (2)
k 6−1
Kontsevitch’s formula giving a star product from a Poisson bracket on an affine

space also works without any modification in the complex case (i.e., 6 = Cn ×C× ).
But as mentioned above the argument used to go from local to global does not work
for complex manifolds, and in general I do not know if a global star product exists
for a given Poisson bracket, even in the symplectic case, nor do I know what the
classification of such algebras looks like (see, however, [20], where it is shown
that even if E may not exist, the category of sheaves of E-modules is defined up to
equivalence).
In the rest of the paper we investigate a special class of star algebras, i.e.,
those which live on a cotangent bundle 6 = T ∗ X − {0}, X a complex mani-
fold, equipped with its canonical Poisson bracket. In this case there is a canonical
global star algebra, viz. the algebra E of pseudo-differential operators, which is
the “microlocalization” of the sheaf DX of differential operators on X. It is known
and easy (cf. below) that any two star algebras with the same symplectic Poisson
bracket are locally isomorphic, so our algebras are classified by H 1 (B6, Aut E). It
is also interesting to compare these with algebras of differential operators, locally
isomorphic to DX on X hence classified by H 1 (X, Aut D): this is done in the next
three sections.
3. Pseudo-differential Algebras
3.1. E - ALGEBRAS
Let 6 = T ∗ X − {0} be the cotangent bundle (minus the zero section) of a complex
manifold X, equipped with its canonical symplectic structure. The basis is B6 =
6/C× = P X, the projective cotangent bundle. There is a canonical star algebra on
6, viz. the algebra of pseudo-differential operators, microlocalization of the alge-
bra of differential operators on X, whose Poisson bracket is the standard Poisson
bracket of T ∗ X. If we choose local coordinates x = (x1 , . . . , xn ) on X and the
dual cotangent coordinates ξ = (ξ1 , . . . , ξn ) on the fibers, the pseudo-differential

product is given by the Leibniz rule for symbols f, g ∈ O: b
X 1
f ∗g = ∂ α f ∂ α g. (3)
α! ξ x
The patching cocycle is the cocycle defined by changes of coordinates: this is
a cocycle because it does patch together total symbols of differential operators
(locally: polynomials in ξ ), to give the global sheaf DX of differential operators.
We are interested in star algebras on 6 associated to the canonical Poisson
bracket: we will call E-algebra such an algebra.
PROPOSITION 1. Any E-algebra is locally isomorphic to E through an operator
b−× .
P ∈D
This result is well known and we just give an indication of the proof: locally
the pseudo-differential algebra E has (topological) generators xi , ξi satisfying the
canonical relations
[xi , xj ] = [ξi , ξj ] = [ξi , xj ] − δij = 0.
If A is a star algebra with the same Poisson bracket, one can construct by suc-
cessive approximations symbols Xi , 4i with the same principal part as xi , ξi and
satisfying the same canonical relations
[Xi , Xj ]A = [4i , 4j ]A = [4i , Xj ]A − δij = 0.
Now there is a unique isomorphism U : E → A which takes xi to Xi and ξi to 4i
b−× .
and this is always a differential operator U ∈ D
Remark 3. The construction also works globally over any open subcone U ⊂
T ∗ Cn which is Stein and contractible (e.g., the set {ξi 6= 0} ⊂ T ∗ B, B a ball in
Cn , or a Stein contractible set). Over such a set, any E-algebra A is isomorphic to
E, and any section α of O(m) is the symbol of a section of Am .
Thus one obtains all E-algebras by gluing together models of E over a covering
of 6 by open conic subsets 6i , using automorphisms of E on the intersections. The
following proposition sums up what was said above:
PROPOSITION 2. Star algebras on 6 = T ∗ X − {0} are locally isomorphic to the
pseudo-differential algebra E. The set AlgE of isomorphy classes is canonically
isomorphic to H 1 (P X, Aut E).
Aut E denotes the sheaf of automorphisms of E; the noncommutative cohomol-
ogy H 1 (P X, Aut E) is described below in Section 3.4.
3.2. DIFFERENTIAL OPERATORS AND D - ALGEBRAS
If X is a complex manifold, the sheaf DX of differential operators on X is well

defined. If U is an isomorphism of DX preserving symbols, it fixes the subalgebra
OX ⊂ DX of operators of order 0 (because it fixes symbols and preserves invertible
operators, which are necessarily of order 0). It follows that U is locally an inner
automorphism of the form Int ef (f holomorphic). We have Int ef = Id iff f is
(locally) constant, so the automorphism sheaf is
Aut DX ' OX× /C× ' OX /C. (4)
We will call D-algebra a sheaf of algebras on X locally isomorphic to DX (such
algebras appear in [2] where they are called “twisted algebras of differential op-
erators”). The set AlgD of isomorphic classes of these algebras is canonically
isomorphic to H 1 (X, OX /C).
A D-algebra obviously also defines a star algebra on P X, and it is natural to
compare the two sets AlgD and AlgE .
3.3. AUTOMORPHISMS AND SYMBOLS OF AUTOMORPHISMS
To understand how local E-algebras can be patched together to make global ob-
jects, we have to know what automorphisms of E look like.
Let U ∈ D b−× be an automorphism of E : U preserves symbols and the unit 1, so
P
U − 1 is of degree 6 −1 and the logarithm D = Log U = − n>1 − n1 (1 − U )n is
well defined; it is a derivation of degree 6 −1 of E.
Now if D is a derivation of degree 6 k its symbol δ = σk (D) is a homogeneous
derivation of degree k of the Poisson algebra O, b i.e., a symplectic vector field on
6, homogeneous of degree k. This corresponds, via the symplectic structure of 6,
to a closed differential form α, homogeneous of degree k + 1.
Let ρ denote
P the radial vector field, infinitesimal generator of the action of
C× (ρ = ξj ∂ξj in local coordinates on X, T ∗ X as above): the associated Lie
derivation is Lρ = iρ d + diρ (iρ denotes the interior product) so
diρ α = (k + 1)α.
Hence α is exact (the differential of a homogeneous function) if k + 1 6= 0. If
k + 1 = 0, s = iρ α is locally constant, and α is locally the differential of a
homogeneous function of degree 0 iff s = 0.
By successive approximations, it follows that locally any derivation D of E is
of the form sad(Log P1 ) + ad Q with P1 elliptic of degree 1, Q ∈ E, and any
automorphism of E is locally of the form
U = (Int P1 )s Int Q0 (5)
with P1 elliptic of degree 1, Q0 elliptic of degree 0.? Int P denotes the inner
automorphism a → P aP −1 .
? As usual in the context of pseudo-differential operators, elliptic = invertible.
If U is an automorphism of E, we define its symbol σ (U ) as the closed 1-form

on 6 homogeneous of degree 0 corresponding to the leading term of Log U .
We have σ (U ) = d Log σ (P ) if U = Int P global section of ω (this is a closed
1-form on 6). If σ (U ) = 0 (Log U of degree 6 2) there exists a unique P ∈ E ×
of degree 0 and symbol 1 such that U = Int P . Summing up we have proved:
PROPOSITION 3. There is an exact sequence of sheaves of groups on P X:
0 → E−× → Aut E → ω → 0, (6)
where E−× denotes the multiplicative sheaf of groups on B6 of sections of E of

symbol 1, and ω is the sheaf on P X of closed 1-forms homogeneous of degree 0
on 6.
If A ∈ AlgE ' H 1 (P X, Aut E) its symbol σ (A) ∈ H 1 (P X, ω) is defined as

the image cocycle.
Remark 4. If U is an automorphism of A, it defines a one parameter group

U = exp sLog U, s ∈ C. This is polynomial in s mod An for any n < 0.
s
3.4. NONCOMMUTATIVE COHOMOLOGY CLASSES
In this section we recall the elementary results of noncommutative cohomology

that we will use (for more information see [16]). Let Y be a space and G a sheaf of
groups on Y . We denote H 0 (Y, G) = Γ (Y, G) the set of global sections of G over
Y : this is a group.
We denote H 1 (Y, G) the set of equivalence classes of cocycles
uij ∈ Γ (Yij = Yi ∩ Yj , G) such that uij uj k = uik

S
associated to open coverings Y = Yi ; two cocycles are equivalent if, after a
suitable refinement of the covering, we have uij = ui u0ij u−1
j for some family ui ∈
Γ (Yi , G).
H 1 (Y, G) classifies the set of isomorphic classes of G principal homogeneous
right G sheaves, i.e., sheaves α on Y , equipped with a right action of G, locally
isomorphic to G considered as a right G-sheaf.
PROPOSITION 4. Let
0→A→B→C→0 (7)
be an exact sequence of sheaves of groups on Y , with A normal in B. Then there is

a long cohomology sequence;
0 → H 0 (Y, A) → H 0 (Y, B) → H 0 (Y, C)
(8)
→ H 1 (Y, A) → H 1 (Y, B) → H 1 (Y, C).
This is “exact” in the sense that

(i) it is exact at the first three places (the H 0 are groups, the H 1 are pointed
sets).
(ii) The group H 0 (Y, C) acts on the set H 1 (Y, A), and its orbits are the fibers
of the map H 1 (Y, A) → H 1 (Y, B) (the action is given by c · (aij ) = (bi aij bj−1 ) if
c is a global section of B, and bi ∈ Γ (Yi , B) a lifting of c to B over a fine enough
covering Yi ).
(iii) If β ∈ H 1 (Y, B) it defines twisted sheaves of groups Aβ ⊂ Bβ (where Bβ
is the sheaf of B-automorphisms of the principal B-sheaf β), and the fiber of the
map H 1 (Y, B) → H 1 (Y, C) is the image of H 1 (Y, Aβ ) in H 1 (Y, C).
More explicitly, if β, β 0 are two principal B-sheaves, then γ = HomB (β, β 0 )
is a principal Bβ -sheaf. If β, β 0 have the same image in H 1 (Y, C) then γ /Aβ =
HomC (β/A, β 0 /A) has a global section, i.e., is trivial, so γ is the image of a sheaf
α ∈ H 1 (Y, Aβ ). Finally, β 0 ∼ α ×Aβ β is in the image of H 1 (Y, Aβ ).
In this paper, the noncommutative cohomology sequence stops there, and we
will not use higher cohomology H j , j > 2, whose definition is more elaborate
(the substitutes are more complicated objects sometimes described by means of
“stacks”). Exact sequences concerning torsors as above were introduced by Frenkel
[12, 13]. Of course if A, B, C are commutative, the higher cohomology groups
H j , j > 0, are well-defined commutative groups, and we will occasionally use the
long cohomology exact sequence in that case up to order j = 2.
3.5. SYMBOLS
If A ∈ AlgE ' H 1 (P X, Aut E) we have defined its symbol as the image of its
defining cocycle in H 1 (P X, ω). To compute H 0 and H 1 for automorphisms, it
will be useful to compute them first for symbols.
The following exact sequences of sheaves are also useful to handle ω:
0 → OP X /C → ω → C → 0, (9)
0 → C → OP X → OP X /C → 0. (10)
These give rise to long exact cohomology sequences. We will call “exponent
map” the cohomology maps coming from the map ω → C in (9).
With slight abuse we will call “Chern maps”? the maps:
ch : H j (Y, O/C) → H j +1 (Y, C) (11)
in the long exact cohomology sequence derived from (10).
The sheaf O/C (Y = X or P X) identifies with the sheaf of closed holomorphic
1-forms on Y . If Y is a Stein manifold we have H j (Y, O) = 0 for j > 1 so the
Chern map H j (Y, O/C) → H j +1 (Y, C) is an isomorphism for j > 1.
? The standard Chern map: H 1 (Y, O × ) → H 2 (Y, C) factors through H 1 (Y, O/C).
If Y is a compact Kähler manifold, the long exact cohomology sequence from

(10) splits into a sequence of short split exact sequences:
0 → H j −1 (Y, O/C) → H j (Y, C) → H j (Y, O) → 0 (j > 0)
and for j > 0 we have an isomorphism
X
H j (Y, O/C) = H pq , (12)
p+q=j +1,p>0
where (here, and whenever possible) H pq denotes the space of harmonic forms of
type p, q on Y .
PROPOSITION 5. (i) If n = dim X > 2, or if X is a closed curve of genus 6= 1,
the map H 0 (X, O/C) → H 0 (P X, ω) is an isomorphism.
(ii) If X is an open curve or a closed curve of genus 1, then w is split and
H 0 (P X, ω) ' H 0 (X, O/C) ⊕ H 0 (X, C).
Proof. A global section of ω is a closed 1-form on T ∗ X − {0}, homogeneous of
degree 0. Locally on X such a form α reads
X
α= αk dxk + βk dξk , (13)
where the coefficients αk resp. βk are of degree 0 resp. −1. If n > 2 this implies
βk = 0 so the αk only depend on x. Hence (i) for n > 2.
If X is a closed curve of genus 6= 1 (n = 1 so P X = X), then the Chern map
H 0 (X, C) ' C → H 1 (X, O/C) = C is injective: it maps s ∈ C to s ch O(1)
(where as above O(1) denotes the sheaf of homogeneous functions of degree 1 on
T ∗ X) and ch O(1) 6= 0 if g 6= 1.? So the exponent map H 0 (X, ω) → H 0 (X, C)
vanishes, and the map H 0 (X, O/C) → H 0 (X, ω) is an isomorphism, hence (i) in
this case.
If n = 1 and X is open or of genus 1, there exists a global nonvanishing vector
field, so ω is split: ω = O/C ⊕ C, hence (ii). 2
PROPOSITION 6. (i) If n = dim X > 2 the map H 1 (X, O/C) → H 1 (P X, ω) is

an isomorphism.
(ii) If n = dim X = 1 (P X = X) and X is open or closed of genus 1 (ω split),
then H 1 (X, ω) = H 1 (X, O/C) ⊕ H 1 (X, C).
(iii) If X is a closed curved of genus g 6= 1 the exponent map H 1 (X, ω) →
H 1 (X, C) ' C2g is an isomorphism.
This should be complemented as follows in case (ii): if X is an open curve,

H 1 (X, O/C) = 0 so H 1 (X, ω) ' H 1 (X, C).
? The corresponding cocycle is dLog(ξ /ξ ) if ξ is the symbol of a nonvanishing vector field on
i j i
a covering Xi of X, whose image in H 1 (X, ω) is dξi /ξi − dξj /ξj , obviously a coboundary.
If X is closed of genus 1, then H 20 = 0 so H 1 (X, O/C) ' H 20 + H 11 '

H ' C, and H 1 (X, w) ' H 11 + H 1 (X, C) ' C3 .
11
LEMMA 1. If X is a ball (or more generally Stein contractible space), we have

H 1 (P X, ω) = 0.
Proof. We have P X ' X ×Pn−1 , so H 1 (P X, C) = 0 (P X is simply connected)
and the map H 1 (P X, O/C) → H 1 (P X, ω) is onto (if n = 1 we are finished).
Next we wave H j (P X, O) = 0 for any j > 0 (O has no cohomology on Pn−1 )
so the Chern map H 1 (P X, O/C) → H 2 (P X, C) ' C is one to one. Now, as above
for curves of genus 6= 1, H 2 (P X, C) ' C is generated by the Chern class of O(1),
corresponding to the cocycle dLog( ξξji ) for ξi an elliptic symbol of degree 1 over
a covering Ui of P X. This is also precisely the image of 1 ∈ H 0 (P X, C) ' C
by the exponent map H 0 (P X, C) → H 1 (P X, O/C), so the exponent map is onto
and the map H 1 (P X, O) → H 1 (P X, ω) vanishes. This proves the lemma. 2
Proof of Proposition 6. (i) Let α beS

a principal ω-sheaf on P X corresponding to
a cocycle in H 1 (P X, ω), and let X = Xi be a covering of X by complex balls (or
Stein contractible open sets). Then αi = α|Xi is trivial. The patching isomorphism
uij : αj → αi is the translation by a section uij ∈ H 0 (P Xi ∩ P Xj , ω) = H 0 (Xi ∩
Xj , O/C); thus α is defined by a cocycle (uij ) ∈ H 1 (X, O/C). If n > 2 and if
(uij ) = (αi −αj ) ∼ 0 in H 1 (P X, ω) then again αi ∈ H 0 (Xi , O/C) by Proposition
5, so (uij ) ∼ 0 in H 1 (X, O/C). This proves (i).
If n = 1 (P X = X) and X is open or of genus 1, ω is split so H 1 (X, ω) =
H (X, O/C) ⊕ H 1 (X, C).
1
If X is open then H 1 (X, O/C) = 0 because in the long exact cohomology se-
quence from (10) we have H 1 (X, O) = H 2 (X, C) = 0, so H 1 (X, ω) ' H 1 (X, C).
If X is of genus g = 1, we have H 1 (X, O/C) = H 20 + H 11 = C and
H (X, ω) ' H 11 + H 1 (X, C) ' C3 .
1
If X is a closed curve of genus g 6= 1 we have seen that the map H 0 (X, C) →

H 1 (X, O/C) is one to one, and H 2 (X, O/C) = H 30 + H 21 + H 12 = 0 so from
the long exact cohomology sequence from (9)
· · · → H 0 (X, C) → H 1 (X, O/C) → H 1 (X, ω) → H 2 (X, O/C) → · · ·
we see that the map H 1 (X, ω) → H 1 (X, C) is one to one.
This proves Proposition 6 and its complement. Note that if X is a curve, the
only case where H 1 (X, ω) = 0 is when X is simply connected. 2
3.6. FILTRATIONS
As mentioned above Aut E has a natural filtration (as well as E−× ⊂ Aut E): any
a ∈ Aut E is of degree 6 0 and a ∈ E−× is of degree n < 0 if a = Int (1+b), b ∈ En .
The corresponding graded sheaf is
M M
gr Aut E = (Aut E)k /(Aut E)k−1 ' ω + O(k). (14)
k 60 k<0
It is commutative, and this will help to extract more information. This also works
for any E-algebra A because the filtration above, and the leading terms, are by
definition invariant by automorphisms so gr Aut A ' gr Aut E.
PROPOSITION 7. Let A be an E-algebra.

(i) The natural map gr H 0 (P X, Aut A) → H 0 (P X, gr Aut E) is injective. If
H 1 (P X, gr Aut E) = 0 it is one to one.
(ii) If H 1 (P X, gr Aut E) = 0 then H 1 (P X, A×
− ) = 0.
(iii) If H 2 (P X, E−× ) = 0 the symbol map induces a surjective map
H 1 (P X, gr Aut E) → gr H 1 (P X, Aut A). (15)
Proof. (i) The map gr H 0 (P X, Aut A) → H 0 (P X, gr Aut E) takes any U of
degree m 6 0 to its symbol σm (U ) which is a section of ω if m = 0 or of O(m)
if m < 0. σm (U ) = 0 means that U is really of degree 6 m − 1 so the resulting
graded map is injective.
Conversely let Xi be a covering of P X, and ai ∈ H 0 (Xi , Aut A) be such
that ai aj−1 is of degree m < 0 (this is true if σ (ai ) = a, a global section of
grm+1 (P X, Aut E)). Then σm (ai aj−1 ) is a 1-cocycle with coefficients in O(m). If
H 1 (P X, E−× ) = 0 this is a coboundary, i.e., of the form σm (bi ) − σm (bj ), bi ∈ Am
so the Int(1 + bi )−1 ai Int(1 + bj ) are equal to ai mod Aut Am and patch together
mod(Aut A)m−1 . Note that if the Xi and their intersections are Stein it is not nec-
essary to shrink the covering, so by successive approximations we get a cocycle
a ∈ H 0 (P X; Aut A) equal to (ai ) mod(Aut A)m.
(ii) If H 1 (P X, gr E−× ) = 0 and if (aij ) is a cocycle of degree n < 0 with
coefficients in A× − , then σn (aij ) is a 1-cocycle with coefficients in O(n), hence a
coboundary σn (bi ) − σn (bj ), bi ∈ An . So a is equivalent to the cocycle
(1 + bi )−1 aij (1 + bj )
which is of degree n − 1. Again we do not need to shrink the covering if it has

been chosen as above (Stein, contractible), so by successive approximations we
get a ∼ 0.
(iii) If bij is a cocycle with coefficients in Autm A (m 6 0) its symbol σ (bij ) is
a cocycle with coefficients in grm Aut E but in general this does not give rise to a
map gr H 1 (P X, Aut A) → H 1 (P X, gr Aut E) nor the other way round, at best an
ill-defined “noncommutative spectral sequence”.
However, if H 2 (P X, gr Aut E = 0), the same argument as above shows that
if a cochain aij ∈ H 0 (Ui ∩ Uj , Aut A) is a cocycle mod(Aut P A)m , m < 0, i.e.,
aij aj k aki ∈ (Aut A)m then σm (aij aj k aki ) is a coboundary σm (bj k ) with coeffi-
cients in O(m), and again by successive approximations there exists a cocycle aij0
with coefficients in Aut A equal to aij mod Am .
In particular, by successive approximations, we see that any cocycle a with
coefficients in grm Aut E (m 6 0) is the symbol of a cocycle b ∈ H 1 (P X, Autm A)
which is well-defined mod H 1 (P X, Autm−1 A) and vanishes if a is a coboundary.
Thus our map is well defined and onto (if H 0 (P X, Aut A) 6= 0 it may not be
injective because two cocycles of degree m with coefficients in Aut A can then be
equivalent although their symbols are not). 2
4. E -algebras on T ∗ X, dim X > 2

4.1. GENERAL RESULTS
We first point out the following results (which will also be useful in Section 5):
LEMMA 2 (Global automorphisms of E). If A is an E-algebra A on X and n =

dim X > 2 the symbol map
H 0 (P X, Aut A) ' H 0 (P X, ω) ' H 0 (X, OX /C)
is injective. It is bijective if A = E.
Proof. If n > 2, A and Aut A have no global section of degree < 0 so the sym-
bol map u → σ (u) is injective. More generally, if A, A0 are two E-algebras and
u, v two isomorphisms A → A0 the difference symbol σ (u−1 v) ∈ H 0 (P X, ω) is
well defined and completely determines v (given u) (note that we have σ (u−1 v) =
σ (vu−1 )).
On the other hand, if A = E (more generally if A comes from a D-algebra)
the symbol map is onto because, by Proposition 5, H 0 (P X, ω) ' H 0 (X, O/C) '
Aut D, and this obviously lifts to Aut E.
If X is a ball of Cn or more generally a Stein contractible domain, we have
H 1 (P X, ω) = 0 (Lemma 1) so H 1 (P X, E−× ) → H 1 (P X, Aut E) is onto, i.e., any
E-algebra can be defined by a cocycle with coefficients in E−× .
Now let A, A0 be two algebras defined by cocycles a = (aij ), a 0 = (aij0 ) ∈ E−×
and let u : A0 → A be an isomorphism, i.e., a family (ui ) ∈ Aut E such that
ui aij0 = aij uj . Then the symbols σ (ui ) patch together since σ (aij ) = σ (aij0 ) = 0,
and the resulting symbol σaa 0 (u) is well defined. It only depends on the classes of
a, a 0 in H 1 (P X, E−× ). However, it does depend on a, a 0 ∈ H 1 (P X, E−× ) and not
just on their images in H 1 (P X, Aut E): any other representatives are of the form
α · a, α 0 · a 0 with α, α 0 ∈ H 0 (P X, ω) for the action of H 0 (P X, ω) of Proposition 4,
and we get σα·a,α 0·a 0 (u) = σ (u) + α − α 0 .
If X ⊂ Cn , n > 2, is a Stein contractible domain, the exponent of σaa 0 (u)
vanishes: σaa 0 (u) ∈ H 0 (X, O/C), and again σaa 0 (u) completely
S determines u.
Let now A ∈ AlgE . There exists a covering X = Xi where all finite inter-
sections are isomorphic to Stein contractible domains of Cn . Then Ai = A|Xi can
be defined by a cocycle (ai ) with coefficients in E−× ; this being so the patching iso-
morphisms uij all have exponent 0 and are determined by their symbols σai aj (uij )
(for fixed Ai ). In particular we have proved:
PROPOSITION 8. If dim X > 2 any E-algebra A has exponent 0 (the image

of σ (A) ∈ H 1 (P X, ω) in H 1 (P X, C) by the exponent map is zero), so A can
be defined by a cocycle with coefficients in Int E0 = E0× /C× (E0 is the sheaf of
pseudo-differential operators of degree 6 0).
4.2. THE CASE dim X > 3
If X is a ball and dim X > 3 we have H 1 (P X, O(−k)) = 0 for all k > 0, i.e.,
H 1 (P X, gr E−× ) = 0 (this is also true if X is a Stein manifold).
It follows that we have H 1 (P X, E−× ) = 0, and more generally for any E-algebra
A we have H 1 (P X, A× − ) = 0.
Hence if A is an E-algebra, it is built by patching together models of E over a
covering Xi of X, where the patching cocycle belongs to H 1 (X, OX /C).
Moreover, if A, B are two such algebras, any isomorphism B → A comes
from a ϕ ∈ H 0 (X, OX /C), i.e., comes locally from an inner automorphism P →
ϕP ϕ −1 . Summing up we have proved:
THEOREM 1. If dim X > 3 the functor which to a D-algebra associates the

corresponding E-algebra is an equivalence.
This result is closely related to the result of [8] on microlocally free D-modules
in dimension > 3.
4.3. THE CASE dim X = 2
If dim X = 2 what was said above remains true, in particular any symbol α ∈
H 1 (P X, ω) is the symbol of an E-algebra (in fact of a D-algebra). However,
the picture changes considerably because H 1 (P X, E−× ) is usually very large. The
following examples show what can happen, and also how, in global situations on
compact manifolds, things can nevertheless at least partially cancel out.
EXAMPLE 1. Let X be the unit ball of C2 (or more generally a Stein contractible
manifold).?
Then H 1 (X, ω) = H 1 (X, O/C) = 0 so H 1 (P X, Aut E) is the quotient of
H 1 (P X, E−× ) by the action of H 0 (X, Aut E) = H 0 (X, O/C).
Now P X is the union of the two Stein subcones Ui = {ξi 6= 0} (i = 1, 2) so a
cocycle is represented by just one section a12 ∈ E−× (U1 ∩ U2 ). It is elementary that
any a ∈ H 1 (P X, E−× ) has a unique normalized representative of the form
X p q
a12 = apq (x)ξ1 ξ2 , (16)
p,q<0
? What is used is H 1 (X, O/C) = 0 and the fact that T ∗ X is a trivial holomorphic vector bundle.
i.e., with no holomorphic term in ξ1 or ξ2 (this is obvious for the additive co-
homology H 1 (P X, gr E−1 ) and follows by successive approximation for E−× ). So
H 1 (P X, Aut E) is the set of conjugate classes of normalized symbols a12 as above,
with a12 ∼ ϕ(x)a12 ϕ(x)−1 for ϕ a nonvanishing function on X. This set is still
very large; on the other hand such algebras tend to have very few global sections
or automorphisms.
The analysis of these algebras is closely related to that of “microlocally” free
D-modules in dimension 2, made by M. Carette [7].
For global compact manifolds, some things may cancel out.
EXAMPLE 2. Let X = P2 (C) be the complex projective plane: then P X is
isomorphic to the incidence manifold {x · ξ = 0} ⊂ X × X∗ (X∗ the dual projective
space). T ∗ X itself is the quotient of the incidence cone 0 = {x · ξ = 0} ⊂
C − {0} × C by the group action (x, ξ ) ∼ (λx, (1/λ)ξ ). The sheaf OP X (n) of
homogeneous functions of degree n on T ∗ X identities with the sheaf of restric-
tions to 0 of functions f (x, ξ ) such that f (λx, ξ ) = f (x, λξ ) = λn f (x, ξ ),
i.e., OP X (n) = OX (n) ⊗ OX∗ (n) (where exceptionally here OX (n) denotes the
canonical sheaf of the projective space). It follows easily that H 1 (P X, gr E−× ) = 0
so H 1 (P X, A×− ) = 0 for any E-algebra A.
The symbol map H 1 (X, Aut E) → H 1 (X, ω) = H 1 (X, O/C) is one to one,
and again, as in dimension > 3, the correspondence D-algebra → E-algebras is
an equivalence.
Note that in this case we have H 1 (P X, O/C) ' H 20 + H 11 ' H 11 = C, and
E-algebras ∼ D-algebras are parameterized by H 11 = C.
EXAMPLE 3. Let X be a holomorphic complex torus of dimension 2 (a torus

C2 / 0 with 0 ' Z4 acting by translations).
The group of automorphisms of E or D is
H 0 (P X, Aut E) = H 0 (X, O/C) = H 10 = C2 (17)
and any automorphism comes from an inner automorphism of E and D on C2 of
the form:
P = P (x, d) → ea.x P e−a.x (x → x, d → d − a). (18)
Any E- or D-algebra on X lifts as the trivial algebra EC2 on the universal cover
2
C , and is the quotient of EC2 by a group of isomorphisms over the translation
group of periods 0.
By Proposition 6 we have
H 1 (P X, ω) = H 1 (X, O/C) = H 20 + H 11 = C5 .
More precisely, an element α ∈ H 20 + H 11 is represented by a harmonic form
X
α = adz1 dz2 + aij d z̄i dzj . (19)
There is a unique corresponding D-algebra, which is isomorphic to the quotient of

DC2 by the lifting µ → Uµ of the group 0 of periods (acting by translations):
x →x+µ
Uµ : (20)
d → d + p(µ, µ),
where p = (p1 , p2 ) is a linear map C4 = C2 × C2 → C2 such that dp(z, z̄) · dz =

dp1 dz1 + dp2 dz2 = α, where z, resp. z̄ denotes the variable in C2 resp. C2 , and we
use the notations of differential calculus. Such a map p splits into holomorphic and
antiholomorphic parts: p = p 0 (z)+p 00(z̄). They form an 8-dimensional space, but it
is classical that maps which differ by a symmetric holomorphic map (dp.dz) = 0)
define isomorphic algebras.
Remark 5. Cocycles coming from H 11 are related to holomorphic line bundles
on X: if L is a line bundle, DL the sheaf of differential operators on L, the corre-
sponding cocycle is the image in H 1 (X, O/C) of the multiplicative cocycle with
coefficients in O × defining L; the corresponding harmonic form is an integral form
in H 11 , and such forms generate H 11 if X is algebraic.
The cocycle associated to dz1 dz2 ∈ H 20 corresponds to the group of isomor-
phisms
Uµ : z → z + µ, d1 → d1 , d2 → d2 + µ1 . (21)
This corresponds to the 1-form p(x) · dx = x1 dx2 (which could be replaced by any
holomorphic primitive of dx1 dx2 ). It never appears in a context of line bundles.
We may now classify E-algebras. The map H 1 (P X, Aut E) → H 1 (P X, ω) is
onto, and for α ∈ H 1 (P X, ω) ' H 20 + H 11 the fiber σ −1 (α) is the image of
H 1 (P X, A×− ) for A the unique D-algebra as above with this symbol.
Let us examine H 1 (P X, A× 1 ×
− ): by (18) two elements of H (P X, A− ) give the
same E-algebra iff there is a translation ξ → ξ + a which transforms one to the
other. An a ∈ H 1 (P X, A× ×
− ) lifts to an element ã ∈ H (C , E− ) invariant by the
1 2
Uµ , so the normalized representative (16) is invariant:

a(x, ξ ) = a(x + µ, ξ + p 0 (µ) + p 00 (µ)), (22)
where as above p 0 : C2 → C2 , resp. p 00 : C2 → C2 denote the holomorphic and
antiholomorphic parts of p, which correspond to the H 20, H 11 components of the
symbol α. Equivalently the symbol b(x, ξ ) = a(x, ξ − p 0 (x)) satisfies
X 1 γ
b(x, ξ ) = b(x + µ, ξ + p 00 (µ)) = ∂ b(x + µ, ξ )(p00 (µ))γ . (23)
γ! ξ
If p 00 = 0 this means that b does not depend on x (it is periodic hence constant).
If p 00 6= 0, the periodicity condition implies b = 0: for if b is of degree n 6 −1,
its leading term is periodic in x hence independent of x: bn = bn (ξ ); the next term
satisfies
bn−1 (x, ξ ) − bn−1 (x + µ, ξ ) = bn0 (ξ ) · p 00 (µ) (24)
so it is linear in x: bn−1 = β(ξ ) + γ (ξ ) · x with p 00 (µ) = −γ (ξ ).µ. Since p 00 is

antiholomorphic this implies bn0 = 0 so bn = 0 since its degree is negative.
Summing up we have proved:
PROPOSITION 9. If X is a torus (C2 / 0, 0 ' Z4 ), we have

H 1 (P X, ω) ' H 1 (X, O/C) ' H 20 (X) + H 11 (X) ' C5 . (25)
Any symbol α ∈ H 1 (P X, ω) is the symbol of a unique D-algebra on X.
If the H 11 component of α is 6= 0 there is no other E-algebra with this symbol.
If α ∈ H 20 the E-algebras with symbol α can be defined by a normalized cocycle
X p q
b(ξ ) = bpq ξ1 ξ2 (bpq ∈ C) (26)
p,q<0
whose coefficients are translation invariant (independant of x). Two such cocycles
b, b0 define the same E-algebra iff b0 (ξ ) ' b(ξ + a) for some constant vector a.
P 1 α α
(This is an asymptotic relation between symbols: b(ξ + a) = ∂ ba .)
α! ξ
5. E -Algebras over Curves (dim X = 1)

We now describe E-algebras, and compare them to D-algebras, when X is a curve
(dim X = 1). In this case P X = X. The general method is the same but as we
will see the classification is strikingly different depending on whether X is an open
curve, or a closed curve of genus g = 0, 1 or > 2.
5.1. OPEN CURVES
If X is an open curve, the exponent map H 1 (X, ω) → H 1 (X, C) is an isomorphism

(Proposition 6). Also X is Stein, so H j (X, O(n)) = 0 for j > 0 and for all
n, j > 1, so H j (X, gr E−× ) = 0 for j = 1, 2, and H 1 (X, E−× ) = 0 (Proposition 7).
Finally we have
H 1 (X, Aut E) ' H 1 (X, ω) ' H 1 (X, C). (27)
Typically, if (sij ) is a cocycle with coefficients in C, the corresponding algebra
is defined by a cocycle with symbol (Int ξ )sij , ξ a global nonvanishing vector field.
These algebras have many sections because we have H 1 (X, E−× ) = 0 so by
Proposition 7 the map H 0 (X, gr E) ' O(X)[ξ, ξ −1 ] → gr H 0 (X, A) is one to one.
They also have many automorphisms, because the sequence 0 → H 0 (X, A× −) →
H 0 (X, Aut A) → H 0 (X, ω) → 0 is exact.
D-algebras are classified by H 1 (X, O/C) = 0 and all give isomorphic E-
algebras. All nontrivial E-algebras come from the exponent map.?
? The fact that such “exotic” algebras exist is related to the fact that coherent D-modules do not
always possess global good filtrations.
5.2. CURVES OF GENUS g > 2
Note that in any case OP X (1) identifies with the sheaf of sections of T X (vector
fields) so the dual OP X (−1) identifies with the sheaf of sections of T ∗ X . If X is
of genus > 2, this is ample so we have H 1 (X, OP X (n)) = 0 for all n < 0, and it
follows that H 1 (X, gr E−× ) = 0, so H 1 (X, A×
− ) = 0 for any E-algebra A, and the
canonical map gr H 0 (X, Aut A) → H 0 (X, gr Aut E) is bijective for any E-algebra
A (it is surjective since on any curve we have H 2 (X, gr E−× ) = 0).
Furthermore, in this case the Chern map H 1 (X, O/C) → H 2 (X, C) is one
to one, as well as the map H 0 (X, C) → H 1 (X, O/C) (cf. 9). Finally the map
H 1 (X, Aut E) → H 1 (X, ω) = H 1 (X, O/C) is one to one. Thus
PROPOSITION 10. If X is a closed curve of genus g > 1, E-algebras on X are

classified by their exponent σ (A) ∈ H 1 (X, C) = C2g . D-algebras are classified
by H 1 (X, O/C) = H 1 (X, C) = C and give isomorphic E-algebras.
Here again E-algebras on X have many sections of Lnegative degree because

H 1 (X, E−× ) vanishes and the map H 0 (X, gr E) ' n60 H (X, O(n)) →
0
0
gr H (X, A) is one to one. They also have many automorphisms because the se-
quence 0 → H 0 (X, A× − ) → H (X, Aut A) → H (X, ω) ' C → 0 is exact.
0 0 g
5.3. CURVES OF GENUS 1
This is the most complicated of the cases examined here. Let X be a closed curve
of genus 1: X = C/ 0 where the group of periods 0 ' Z2 acts by translations.
We denote ξ the symbol of the constant vector field ∂/∂x on C.
Since T X is trivial, ω is split: ω = O/C + C. Also, for all n, we have
H 0 (O(n)) = H 00 ' H 1 (X, O(n)) = H 01 ' C, H 2 (O(n)) = 0.
We denote
G, resp. G− ⊂ G (28)
the group of automorphisms of E of the form Int ξ s Int(1 + a(ξ −1 )), resp. the sub-
group s = 0: this is the commutant of ξ , it is a constant subsheaf of Aut E.
For any α ∈ C we set ξa = eax ξ . This is only defined up to a multiplicative
constant eaµ , µ ∈ 0, but the inner automorphism Int (eax ξ ) is well defined, as well
as the corresponding commutator sheaf
Ga− ⊂ Ga , (29)
which is a locally constant subsheaf of Aut E.
PROPOSITION 11. We have
H 0 (X, ω) = H 0 (X, O/C) + H 0 (X, C) = H 10 + H 00 ' C2 ,

H 1 (X, ω) = H 1 (X, O/C) + H 1 (X, C) = H 11 + (H 10 + H 01 ) = C3 .
For the commutative locally constant sheaf Ga− we have?

H j (G− ) = H j (X, C) ⊗ G− if a = 0, 0 if a 6= 0.
L
We have gr E− = n<0 O(n)ξ n and with an obvious notation
H 0 (X, gr E−× ) = gr G ' ξ −1 C[ξ −1 ],
H 1 (X, gr E−× ) ' H 10 × ξ −1 C[ξ −1 ],
H 2 (X, gr E−× ) = 0.
THEOREM 2. If X is of genus 1, the symbol map AlgE → H 1 (X, ω) is onto. We
will denote σ (A) = α = (α 11, α 10 , α 01 ) ∈ H 11 × H 10 × H 01 the symbol of an
E-algebra A. Then
(i) Algebras such that α 11 = 0 are characterized by the fact that they possess a
global section of degree 6= 0, or an automorphism of symbol dξ /ξ = σ (Int ξ ). For
such an algebra the set of global sections is C((ξ −1 )) and except for E the group
of automorphisms is G.
E is distinguished by the fact that its symbol map H 0 (X, Aut E) → H 0 (X, ω)
is onto.
(ii) If α 11 6= 0, A has no section of degree 6= 0 (H 0 (X, A) = C), and A
is completely determined by its symbol, in other words the image of H 1 (X, A) in
H 1 (X, Aut A) is reduced to a single point.
For such an algebra the group of automorphisms is a one parameter group
with symbol C(a dx + b(dξ /ξ )) for some (a, b) 6= 0, except in the case α 01 =
0, α 11 , α 10 6= 0 where there is no automorphism other than Id.
(iii) Among these, E-algebras associated to a nontrivial D-algebra are those for
which σ (A) = α 11 ∈ H 11 (α 10 = α 01 = 0). They are characterized by the fact that
their group of automorphisms is a one parameter group with symbol σ (Int et x ) (t ∈
C).
Thus, for a torus X of genus 1, D-algebras which give isomorphic E-algebras
are already isomorphic as D-algebras, and E-automorphisms are the same as D-
automorphisms, except for the canonical algebra E.
Let A be an E-algebra. The symbol map AlgE → H 1 (X, ω) is onto because
H 2 (X, gr E−× ) = 0 so (Proposition 7) any cocycle with coefficients in ω is the
symbol of an E-algebra.
Next note that any star algebra on X lifts as the trivial E-algebra EC on C with
an action of the group 0 over the translation group:
µ → Tµ = Tµ Uµ , (30)
where Tµ is the translation (x → x + µ, ξ → ξ ) and the Uµ are automorphisms
of EC , subjected to the cocycle condition expressing that µ → Tµ Uµ is a group
homomorphism.
? The cohomology of the locally constant sheaf generated by eax ξ or enax ξ n vanishes if na 6= 0,
because eanµ cannot be identically 1 for µ ∈ 0, so H ∗ (X, gr Ga− ) = 0.
Here are typical examples (models):
EXAMPLE 4. Let µ ∈ 0 → Uµ = α(ξ ) ∈ G be an additive map. This defines

such a cocycle, because the Uµ commute with translations, hence an E-algebra,
obviously of the first type since ξ is invariant. Typically the period group
10 µ+α 01 µ
µ → Uµ = (Int ξ )α
defines such an E-algebra with symbol α 10 + α 01 (α 11 = 0).
EXAMPLE 5. Let α(µ) = α 10 µ + α 01 µ be an additive map 0 → C and a ∈ C.

Then the automorphisms Int(exp α(µ)(ax + Log ξ )) commute and also commute
with translations (because the commutator [ξ, ax + Log ξ ] = a is a constant so
exp s(ax + Log ξ ) commutes with translations, mod constant factors which give
trivial inner automorphisms). So the group homomorphism
µ → Uµ = exp α(µ)(ax + Log ξ )
defines an E-algebra, whose symbol is aα 01 , α 10 , α 01 ∈ H 11 × H 10 × H 01 . If we

identify the H pq with spaces of differential forms, the symbol of A writes
σ (A) = (a dα(x, x̄) d x̄, dα(x, x̄)).
Such an algebra admits the automorphisms Int exp s(ax + Log ξ ), s ∈ C.

The only symbols we have missed are those for which
a = α 11 6= 0, b = α 10 6= 0, α 01 = 0.
As model for this case we can take the algebra defined by
Uµ = (Int ξ )bµ Int(eaµ x̃ )
with x̃ = x(1 + b/ξ )−1 so that σ (x̃) = x, [ξ + bLog ξ, x̃] = 1: with this choice the
Tµ (Int ξ )bµ (symbolically exp µ(ξ + bLog ξ )) commute with the Int eaµ x̃ , so again
the Uµ define an E-algebra with symbol α 11 = a, α 10 = b, α 01 = 0.
We now prove Theorem 2.

(i) First suppose that A has a nonzero section s of degree 6= 0. Then σ (s) = cξ k
for some constant c 6= 0 and integer k 6= 0; c−1 s has a unique kth root with symbol
ξ (this is true locally because it works for pseudo-differential calculus; the roots
with symbol ξ are unique and patch together into a global section). Similarly if a
is an automorphism with symbol σ (Int ξ ), locally there exists a unique section s
with symbol ξ such that a = Int s (a = Int b determines b locally up to a constant,
and σ (b) = ξ fixes the constant so again these patch into a global section with
symbol ξ ).
If A has
P a section s with symbol ξ , then clearly all global sections of A are of
the form k6k0 ck s k , and H 0 (X, Aut A) contains the group G (formula (28)).
Furthermore, again by elementary pseudo-differential calculus, any two sec-

tions, resp. automorphisms of E of symbol ξ are locally conjugate, so (A, a) is
locally isomorphic to (E, Int eξ ), and A can be defined by a cocycle with coeffi-
cients in G, the commutator of Int ξ (formula (28)). Hence α = σ (A) belongs to
the image of H 1 (X, G) in H 1 (X, ω), i.e., α ∈ H 1 (X, C) and α 11 = 0.
Conversely let α ∈ H 1 (X, C) (α 11 = 0). Example 4 gives an algebra A with
symbol α which has a section of symbol ξ .
Now any other algebra A0 with symbol α is defined by a cocycle (aij ) ∈
H (X, A×
1 1
− ). We know that there is a surjective map from H (X, gr A− ) to
×
gr H (X, A− ) and also that the map H (X, G) → H (X, E− ) = H (X, A×

1 × 1 1 × 1
−)
is surjective. It follows that the embedding G− → A× − gives a surjective map
H 1 (X, G) → H 1 (X, A× − ) (the symbol map (gr) is onto, and surjective follows
by successive approximations). Thus any E-algebra with symbol α can be defined
from A by a cocycle with coefficients in G− , or from E with coefficients in G; in
particular it has a section with symbol ξ .
LEMMA 3. If such an algebra A (α 11 = 0) is not trivial, it has no other global
automorphism than those of G.
Proof. If A has two automorphisms a, b with independent symbols, we may
suppose that these symbols are σ (Int ξ ), σ (Int ex ). So A has a section α with
symbol ξ such that a = Int α. The section Log b = β is locally well defined up to
an additive constant, so the section γ = [α, β] is globally defined and commutes
with α (as any global section of degree < 0).
Now the symbol γ is 1, so γ is invertible, and replacing β by βγ −1 , we see
that we can suppose [α, β] = 1 (or equivalently b−1 ξ b = ξ + 1). It follows again,
by successive approximations, that A equipped with two such automorphisms is
locally isomorphic to E equipped with Int ξ and Int ex ; but the only automor-
phisms which commute with both are obviously trivial (the leading term is constant
because it commutes with x and ξ ), so A is isomorphic to E.
(ii) Suppose now α 11 6= 0. Then any section is a constant (of degree 0) and
there is no global section of degree 6= 0. Let us choose an algebra A with symbol
α (one of the models above). Here again since H 2 (X, gr E−× ) = 0 the graded map
H 1 (X, gr A× ×
− ) → gr H (X, A− ) is surjective, and any cochain with coefficients in
1
A×− which is a cocycle mod An (n < 0) is equivalent mod An to a cocycle. 2
LEMMA 4. We have H 1 (X, A× − ) = C.
S Proof. Let A be defined by a cocycle Uij relative to some Stein covering X =

Xi . We can choose Uij = Int eαij x Int ξ sij mod lower order terms, with constant
αij , sij .
Then α 11 = σ (A)11 ∈ H 11 corresponds to the (01) part of the cocycle αij , and
does not vanish. Let a = (ai ) be the “constant” 1-cochain ai = ξ n (n 6 0) on Xi .
Its coboundary with coefficients in A× − (computed in A(Xi )) is of degree n − 1;
more precisely on Xi we have
ai − αij (aj ) = ξ n − aij (ξ n ) = ξ n − (ξ − αij )n + · · ·

= nαij ξ n−1 + · · · , (31)
where the . . . are lower order terms, because Int ξ commutes with ξ and the rest
only contributes to terms of degree 6 n − 2.
It follows by successive approximations that any cocycle of degree < −1 is
equivalent to 0, and the same computation shows that two cocycles with the same
leading term are equivalent. Since the symbol map H 1 (X, A×
− ) → H (X, O(−1))
1
× −1
= H is onto, we have H (X, A− ) = C: any cocycle βij ξ is the symbol of a
01 1
unique element of H 1 (X, A× − ). 2
LEMMA 5. The map H 1 (X, A× − ) → H (X, Aut A) is constant.

1
Proof. The fibers of this map are the orbits of the action of H 0 (X, ω) (Proposi-
tion 4). We will prove that this is transitive.
For this action σ (Int ξ ) acts by u = (uij ) → us = usij with Int usij = Int ξis Int uij
Int ξjs (mod coboundary equivalence), where ξi ∈ A(Xi ) has symbol ξ , and mul-
tiplication is the multiplication of A. Now in the local frame on Xi we have
ξj = Uij (ξ ) = x − αij + · · · so for leading terms we get
ξ
σ (usij ) = σ (uij ) = σ (u)(1 + αij ξ −1 ) (32)
ξ − αij
or with additive notation σ (us ) = σ (u) + sα 11 .
This proves Lemma 4, and the other assertions of Theorem 2 are immediate
consequences. 2
5.4. THE PROJECTIVE LINE
Let X be the projective line (X = P1 (C)). It is the union of the two open sets
X0 = {z = 6 ∞}, X∞ = {z 6= 0}, and since these are Stein, contractible (' C), E-
or D-algebras are classified by cocycles reduced to one function on X0 ∩ X∞ .
D-algebras are classified by H 1 (X, O/C) = H 2 (X, C) = C. The D-algebra
Ds (s ∈ C = H 1 (X, O/C)) is defined by the cocycle (Int z)s .
Let us introduce homogeneous coordinates x, y (z = xy ). We make use of the
sheaf of homogeneous differential operators D hom on C2 , i.e., differential operators
of x and y which commute with the generator of homotheties ρ = x∂x + y∂y ; this
algebra is generated by ρ and the operators
e = x∂y , h = x∂x − y∂y , f = y∂x (33)
which satisfy the relations
[h, e] = 2e, [h, f ] = −2f, [e, f ] = h,
h2 + 2(ef + f e) = ρ(ρ + 2). (34)
Ds is isomorphic to the quotient sheaf D hom/(ρ + s) and can be thought of as

the sheaf of differential operators on the virtual sheaf O(s) on X of homogeneous
functions of degree s of x, y (which really only exists s when s is an integer).
We now turn to E-algebras.
LEMMA 6. If X = P1 is the projective line then
(i) H 0 (X, gr Aut E) = 0 hence H 0 (X, Aut A) = 0 for any E-algebra A.
(ii) H 1 (X, ω) = 0 and H 1 (X, E−× ) → H 1 (X, Aut E) is one to one.
Proof. In homogeneous coordinates as above, O(n) is the sheaf of homoge-
neous functions of degree 2n of x, y and has no global section if n < 0, hence
H 0 (X, gr E−× ) = H 0 (X, E−× ) = 0.
We have proved above (Propositions 5, 6) that for the projective line we have
H (X, ω) = H 1 (X, ω) = 0, hence the lemma.
0
Note that any q ∈ H 1 (X, Aut E) has a unique “normalized” representative:

X
q0∞ = b 0 ∩ X∞ ).
apq zp ζ q ∈ O(X (35)
0>p>2q
This is because the two vector fields ∂0 = ∂z , ∂∞ = ∂1/z are globally holomorphic
and elliptic on X0 , resp. X∞ and their symbols are ζ, −z2 ζ , so any cocycle can
uniquely be reduced to the form above, as for the additive cohomology group
H 1 (X, gr E−× ).
To compare D-algebras and E-algebras it is convenient to use the following
intermediate exact sequence; let Int E0 ' E−× /C× be the group of inner automor-
phisms of E0 ; we have an exact sequence:
0 → Int E0 → Aut E → C → 0
hence a surjection
H 1 (X, Int E0 ) → H 1 (X, Aut E) (36)
whose fibers are the orbits of the action of C = H 0 (X, C) on H 1 (X, Int E) (q0∞ →
(Int ∂0 )s q0∞ (Int ∂∞ )−s , cf. Proposition 4). 2
LEMMA 7. We have the following relation:

(Int z)−s−2 = (Int ∂0 )s+1 (Int z)s (Int ∂∞ )−s−1 . (37)
Proof. If s = k is a positive integer we have

z−k−2 (z2 ∂)k+1 = ∂ k+1 zk .
Indeed both are ordinary differential operators of order k + 1, with leading term
zs ∂ s+1 , which kill all monomials zj , 0 > j > −k.
Identity (37) for arbitrary s follows, because it is polynomial in s mod(Aut E)m ,
for any m < 0.
It follows that Ds resp. D−s−2 give isomorphic E-algebras, although they are
not isomorphic D-algebras. This is the only case where two D-algebras on X = P1
give isomorphic E-algebras: the algebra of global sections is obviously an invariant
of an E-algebra, and in this the global sections e, h, f (with the notations above)
are well defined (up to an additive constant by their symbols, and the commutation
relations fix these constants). It follows that s(s + 2) = h2 + 2(ef + f e) is an
invariant of the E-algebra coming from Ds .
Note that D-algebras form a one-parameter family, so there are many E-algebras
which do not come from an E-algebra.
As last remark we turn to the following problem: does there exist a global
symbolic calculus, i.e., is the underlying sheaf of a given E-algebra isomorphic
b This is always true for real E-algebras, where one can patch global symbols
to O?
using a partition of the unity.
Let us examine what happens on X = P1 (C). There is a canonical 2-covering
of T ∗ X − {0} by C2 − {0}: (u, v) → (z = u/v, ζ = 12 v 2 ). If A is a E-algebra
on 6 = T ∗ X its pull-back on 6 0 = C2 − {0} is a star-algebra for the canonical
Poisson bracket ({v, u} = 1), equipped with an involution above the symmetry
(u, v) → (−u, −v) (note that on 6 0 , u and v are of degree 12 ). If A has a global
b 0 ).
symbolic calculus, its pull-back defines a star-product on O(6
b ) there is (up to isomorphism) only one star algebra law for the
Now on O(6 0
canonical Poisson bracket, generated by u, v with Pthe relation [v,Pu] = 1. Up to

q
isomorphism this is given by the representation apq u v →
p q
apq up ∂u . For
this law there are many global sections (i.e., all polynomials of u, v): the global
sections e, h, f are necessarily
e = − 12 u2 , h = 2u ∗ v + 12 , f = 12 v 2 (38)
because their respective symbols are
σ (e) = −z2 ζ ∼ − 12 u2 , σ (h) = 2zζ ∼ uv, σ (f ) = ζ ∼ 12 v 2 .
These determine e, h, f up to additive constants, and the commutation relations

(34) determine the constants as above.
For these constants we get
h2 + 2(ef + f e) = − 34 so s = − 12 or s = − 32 . (39)
We have proved: 2
PROPOSITION 12. The only D-algebras on P1 for which there is a global to-
tal symbolic calculus are D−1/2 and D−3/2 . In particular there is no global total
symbolic calculus for D.
Acknowledgement
I thank J. J. Sansuc for his kind help in proofreading the manuscript, and the
reviewer for many detailed comments and suggestions.
References
1. Bayen, F., Flato, M., Fronsdal, C., Lichnerowicz, A. and Sternheimer, D.: Deformation theory
and quantization I, II, Ann. Phys. 111 (1977), 61–131.
2. Beilinson, A. and Bernstein, J.: Localisation de g-modules, C.R. Acad. Sci. Paris, Sér. I 292
(1981), 15–18.
3. Boutet de Monvel, L.: On the index of Toeplitz operators of several complex variables, Invent.
Math. 50 (1979), 249–272.
4. Boutet de Monvel, L.: Star products on conic Poisson manifolds of constant rank, Mat. Fiz.
Anal. Geom. (Kharkov) 2(2) (1995), 1–9.
5. Boutet de Monvel, L. and Guillemin, V.: The Spectral Theory of Toeplitz Operators, Ann. of
Math. Stud. 99, Princeton University Press, 1981.
6. Boutet de Monvel, L. and Krée, K.: Pseudodifferential operators and Gevrey classes, Ann. Inst.
Fourier 17 (1967), 295–323.
7. Carette, M.: Exotic D-modules in dimension 2, thèse Paris VI, en préparation.
8. D’Agnolo, A. and Schapira, P.: The Rado–Penrose correspondence II: Line bundles and simple
D-modules, J. Funct. Anal. 153(2) (1998), 343–356.
9. De Wilde, M. and Lecomte, P.: Existence of star products and of formal deformations of the
Poisson Lie algebras of arbitrary symplectic manifolds, Lett. Math. Phys. 7 (1983), 487–496.
10. De Wilde, M. and Lecomte, P.: Formal deformations of the Poisson Lie algebra of a symplectic
manifold and star products. Existence, equivalence, derivations, in M. Hazewinkel and M.
Gerstenhaber (eds), Deformation Theory of Algebras and Structures and Applications, Kluwer
Acad. Publ., Dordrecht, 1988, pp. 897–960.
11. Fedosov, B. V.: A simple geometrical construction of deformation quantization, J. Differential
Geom. 40 (1994), 213–238.
12. Frenkel, J.: Cohomologie à valeurs dans un faisceau non abélien, C.R. Acad. Sci. Paris 240
(1955), 2368–2370.
13. Frenkel, J.: Cohomologie non abélienne et espaces fibrés, Bull. Soc. Math. France 85 (1957),
135–220.
14. Gerstenhaber, M.: On the deformation of rings and algebras, Ann. of Math. 79 (1964), 59–103.
15. Guillemin, V. and Sternberg, S.: Geometrical Asymptotics, Amer. Math. Soc. Surveys 14, Amer.
Math. Soc., Providence, RI, 1977.
16. Giraud, J.: Cohomologie non abélienne, Grudlehren Math. Wis. 179, Springer-Verlag, 1971.
17. Gutt, S.: Equivalence of deformations and associated star products, Lett. Math. Phys. 3 (1979),
297–309.
18. Karasev, M. V. and Maslov, V. P.: Pseudodifferential operators and a canonical operator in
general symplectic manifolds, Math. USSR Izv. 23 (1984), 277–305.
19. Karasev, M. V. and Maslov, V. P.: Nonlinear Poisson Brackets: Geometry and Quantization,
Transl. Math. Monogr. 119, Amer. Math. Soc., Providence, 1993.
20. Kashiwara, M.: Quantization of contact manifolds, Publ. RIMS, Kyoto Univ. 32 (1996), 1–7.
21. Kawai, T. and Kashiwara, M.: On holonomic systems of microdifferential equations III –
systems with regular singularities, Publ. RIMS, Kyoto Univ. 17 (1981), 813–979.
22. Kontsevitch, M.: Deformation quantization of Poisson manifolds, Preprint, I.H.E.S., 1997.
23. Meyer, D.: D-modules et E-modules associés à un opérateur à caractéristiques simples, thèse
Paris VI, 1998.
24. Moyal, J.: Quantum mechanics as a statistical theory, Proc. Cambridge Philos. Soc. 45 (1965),
99–124.
25. Vey, J.: Déformation du crochet de Poisson sur une variété symplectique, Comment. Math.
Helv. 50 (1975), 421–454.
26. Weinstein, A.: Deformation quantization, Séminaire Bourbaki 789, 1994.
141
“Momentum” Tunneling between Tori and the

Splitting of Eigenvalues of the Laplace–Beltrami
Operator on Liouville Surfaces
S. YU. DOBROKHOTOV
Institute for Problems in Mechanics of Russian Academy of Sciences, Prosp. Vernadskogo 101,
Moscow 117526, Russia, e-mail: dobr@ipmnet.ru and Institut de Mathématiques de Jussieu, CNRS
UMR 9994, Laboratoire de Physique mathématique e Géometrie, case 7012, Université Paris 7
Denis Diderot, place Jussieu, F-75251 Paris Cedex 05, France, e-mail: dobro@math.jussieu.fr
A. I. SHAFAREVICH
M. V. Lomonosov Moscow State University, Vorob’evy Gory, Moscow 119899, Russia,
e-mail: shaf@difgeo.math.msu.su
(Received: 3 August 1998; in final form: 5 July 1999)

Abstract. Tunneling in the spectral problem for the Laplace–Beltrami operator on a torus with Liou-
ville metric is considered. The formula for exponential splitting of eigenvalues is obtained. The split-
ting can be expressed in terms of gomology and cohomology classes of a complex Lagrangian mani-
fold. These classes are constructed with the help of the phase flow of a certain gradient-Hamiltonian
vector field on the manifold.
Mathematics Subject Classifications (1991): 53C15, 53B20.
Key words: complex Lagrangian manifold, tunneling, Laplace–Beltrami operator, Liouville tori,
semiclassical approximation.
1. Formulation of the Problem and Explanation of Results

1.1. Semiclassical approximation relates certain subsequences of asymptotic eigen-
values (“spectral series”) of Schrödinger-type operators (we denote them by H b)
defined on a Riemannian manifold Q, with families of the invariant tori of the
corresponding classical Hamiltonian systems in the phase space T ∗ Q. For instance,
if the classical Hamiltonian system defined by the Hamiltonian H on T ∗ Q is
integrable in the Liouville sense, and if its motion is on the tori 3(I1 , . . . , In )
parameterized by the action variables I = (I1 , . . . , In ) and belonging to the energy
level surface
E = H |3 = H (I1, . . . , In ), (1.1)
then the “asymptotic eigenvalues” Eν associated with the family 3 are defined by
the Bohr–Sommerfeld rule (see [28, 24]),
142 S. YU. DOBROKHOTOV AND A. I. SHAFAREVICH

σj
Ij = Ijν ≡ h + νj , (1.2)
4
and by the formula (see, e.g., [28, 24, 15, 7, 22])
Eν = H (I1ν , . . . , Inν ). (1.3)
Here, h > 0 is a small parameter (which must be specially interpreted in each
concrete situation), σj is the Maslov index of the cycle associated with the ac-
tion variable Ij , νj are integer, νj ∼ 1/ h. The Maslov canonical operator (or
Fourier integral operator, see, e.g., [24]) also allows one to construct the asymptotic
eigenfunctions (known also as quasimodes) 9ν , approximately (in some norm)
satisfying the original quantum equation
b9ν = Eν 9ν + O(hα ).
H (1.4)
The number α > 1 characterizes the precision of the approximation. The functions
9ν are localized in the neighborhood of the tori 3(I1ν , . . . , Inν ): supp 9ν → π(3)
as h → 0. Here π : T ∗ Q → Q is the natural projection. If H b is a selfadjoint
operator, then in the o(h)-neighborhood of Eν there always exists a number E e
from the spectrum of the operator H b. Therefore if the spectrum of the operator
Hb is discrete, then Eν is an approximation of some exact eigenvalue E. e Generally
speaking, this is not so for the functions 9ν : they may be close to (sometimes rather
complicated) combinations of true eigenfunctions of H b associated with different
eigenvalues (see, e.g., [2, 28]). Some examples of this type are connected with
a situation in which the Hamiltonian H and additional motion integrals of the
corresponding dynamic system are invariant with respect to a certain finite group.
In this case, the asymptotic eigenvalues Eν corresponding to some torus 3 and to
all images of 3 with respect to the group action coincide for all possible α. So we
can see the “asymptotic degeneration” of this part of the spectrum.
The simplest example of such a situation appears in the famous double-well
problem for the Schrödinger equation. In the one-dimensional classical problem
(see [22], and also [15, 39]), there exist two families of invariant circles asso-
ciated with each well, if the energy level is under the barrier between the wells
(the reflection of the coordinate x → −x maps one family onto the other). Each
asymptotic eigenvalue corresponds to two different asymptotic eigenfunctions 9ν±
(each localized in a certain well) for an arbitrarily large α. This “degeneration”
vanishes thanks to tunneling effects between the wells. This effect leads to the
appearance of an exponentially (with respect to the Plank constant h) small splitting
between eigenvalues, and also to the phenomenon that the true eigenfunctions 9 eν±
are localized at once in both wells:
1
eν± = √ (9ν+ ± 9ν− ) + O(h).
9 (1.5)
2
Of course, these effects exist in the multidimensional case (see, e.g., [15, 39, 40,
6]). However, rigorous mathematical investigations have been caried out at only
“MOMENTUM” TUNNELING BETWEEN TORI AND THE SPLITTING OF EIGENVALUES 143
for the lowest energy levels. In this case, we have two equilibrium points instead
of invariant tori, and the splitting of the asymptotic eigenvalues is defined by the
special trajectories of the corresponding Hamiltonian system called instantons. The
integral along the instanton defines something like a “tunnel distance” between
these 0-dimensional invariant sets (see [1, 8, 16, 25, 34 – 36] for the theoretical
treatment, and also [9, 15, 20, 35, 39, 40]).
The main difficulties in the multidimensional case for higher energy levels
are very similar to the difficulties appearing in the multidimensional saddle point
method. In an analysis of one-dimensional tunnel effects so-called canonical paths,
and Stokes and of Stokes lines in 1-D complex configuration space (on the complex
plane) play important roles. The Stokes and anti-Stokes lines are a zero-set of the
imaginary or real part of a certain complex function S(z); the canonical paths are
ones where Im S(z) does not decrease, and this means that it is possible to introduce
the orientation of these paths. Roughly speaking, one can use these objects and the
complex WKB method without knowing the classical (Hamilton) mechanics, the
phase space, invariant tori (circles) in the phase space, etc. But it is well known
that the passing from one-dimensional to multidimensional semiclassical meth-
ods (without tunnel effects) required to lift many objects from the configuration
space to the phase space and recall Hamilton mechanics. This is not trivial for
the construction of quasimodes associated with tori in real phase space (see, e.g.,
[15, 24, 25, 28]) and it is clear that it is much more complicated if one wants to take
into account the tunnel effects in complex phase space.? Also it is unclear how to
generalize the Stokes lines and canonical paths is the multidimensional situation.
That is why we think that it is interesting to consider such a multidimensional
example, where one can rigorously prove the formulas for quasimodes and splitting
connected with the complex phase space and interpret these formulas via some
multidimensional analogs of Stokes and anti-Stokes lines, paths (on the complex
Lagrangian manifold) in complex phase space, etc. This example must differ from
the double well problems, where the final formulas for splitting include integrals
over paths in real (not complex) configuration space (see [22, 15, 39]). Ya. G. Sinai
drew the attention of the first author (S. Yu. Dobrokhotov) to one of such examples
with additional discrete symmetry (see [18]). This is the case of the Laplace–
Beltrami operator h2 1 on a 2-D Liouville torus Q. This example admits separation
of variables and is well studied (without tunneling), in particular in the paper [18]
(see also the references in this paper). The corresponding classical Hamiltonian is
quadratic in momenta and it is obvious that if a torus 3+ = (p = P (ϕ), x = X(ϕ))
is invariant with respect to the phase flow gHt , then the “inverse in momentum”
torus 3− = (p = −P (ϕ), x = X(ϕ)) is also invariant. We denote by (p, x)
the coordinates in the phase space T ∗ Q (x = (x1 , . . . , xn ) are coordinates on Q),
? One of the explanations of this fact is connected with the complicated structure of related geo-
metrical objects. Even in the simplest examples of 1-D (of course integrable) situations, the analogs
of the Liouville tori (circles) are not tori in four real-dimensional space and have a more complicated
geometrical and topological structure.
and by ϕ = (ϕ1 , . . . , ϕn ) the angle coordinates on the tori. If the tori 3+ and 3−
do not coincide, then one can construct two families of quasimodes 9ν+ and 9ν−
associated with these families of tori, which for all possible α in (1.4) correspond
to the same asymptotic eigenvalues Eν . From the point of view of Equation (1.4),
this means that, together with the asymptotic solution 9ν+ , the complex conjugated
+
function 9ν− = 9 ν is also an asymptotic solution of (1.4).?
Now the following natural question arises:
Does splitting of the corresponding asymptotic eigenvalues exist?
Another interesting problem related to this question is:
Is it possible to construct the modes (more exactly, the asymptotics of the real
eigenfunctions; see [2]) using the quasimodes? In particular, is it possible to use
formula (1.5)?
We answer these questions under some additional assumptions, and we not
only give analytical formulas for the splitting of the eigenvalues, and “modes via
quasimodes”, but also their geometrical interpretations. In fact it was not difficult
to obtain these formulas, using separation of variables and combination of some
ideas from [18] and analytical results of the complex WKB method from [13, 33].
Thus, we regard the main result of this paper to be the geometrical interpretation
and our observation that the splitting formula has a natural geometrical and topo-
logical structure. This interpretation includes some multidimensional analogs of
anti-Stokes lines, canonical paths on the complex Lagrangian manifold and also
singularities of this complex manifold and their projections to the related invariant
tori (which are absent in the 1-D case). It shows that it is sometimes possible to
add to the set of geometrical and topological objects (invariant tori 3, the cycles
γj on them, the corresponding action variables Ij and Maslov indices σj ), which
characterize the “spectral series”, certain new geometrical and topological objects
which characterize the splitting and tunneling.
It is important that in the example, considered the classical separation of vari-
ables (we mean angle-actions variables) does not coincide directly with the sepa-
ration of variables in the initial configuration space and the splitting formula really
includes complex paths in the complex phase space (in contrast to examples which
may be reduced to 1-D double-well problems).
The structure of the splitting formula is very transparent. We will describe it
below. Let 3+ and 3− be two “momentum” symmetric invariant tori of the geo-
desic flow on the Liouville torus implying the quasimodes ψν+ , ψν− = ψ̄ν+ and
asymptotic eigenvalues Eν and let there be no other tori with the same energy
and the same value of an additional integral. These tori 3+ and 3− belong to the
real phase space T ∗ Q. Let 8 be the related complex phase space with coordinates
? It is clear that all considerations about “inverse” tori are also true for the Schrödinger-type
operator −h2 1 + V (x) on Q, where V (x) is a potential, and that multidimensional Hamiltonian
systems sometimes can admit larger groups of discrete symmetries and therefore more than two
“identical” families of invariant tori.
z = (z1 , z2 ) ∈C Qz and momenta w = (w1 , w2 ). Then under some additional

assumptions it is possible to construct:
(1) a certain complex (invariant) Lagrangian manifold L in 8 (the complex exten-
sion of tori 3± ) such that L ∩ T ∗ Q = 3+ ∪ 3− ;
(2) a special oriented (“tunnel”) cycle 0 on L connecting 3+ and 3− (the “half”
of the curve 0 going from 3+ to 3− plays the role of instantons in the double
well problems and anti-Stokes lines in 1-D situations);
(3) a special singular (with respect to the projection πz : L →C Q) subset on L
(this subset is a complex caustic) and its certain projection η0 on 3+ . The
projection η0 is a closed curve (two-sided cycle) on 3+ ;
(4) a certain basis γ1 , γ2 of cycles on 3+ ; this basis implies two vectors: a vector
of frequencies ω = (ω1 , ω2 ) of the motion according to the geodesic flow
and the vector m = (m1 , m2 ) with integer components which is a vector of
intersections of (γ1 , γ2 ) and η0 .
The main statement of our paper is that there exist two exact eigenvalues E eν± of
the operator −h 1 on Q such that
2
eν± = Eν + O(h2 ),
E
and
eν+ − E
E eν− = h (ω · m + O(h))e−ρ(3+ , 3− )/ h, (1.6)
π
R
where the “tunnel” distance ρ is equal to (1/2) 0 w dz.
We prove also that formula (1.5) for true asymptotics of eigenfunctions (modes)
via quasimodes holds if the latter are constructed “correctly”. This means the fol-
lowing. Each quasimode ψν± may be multiplied by some complex constant C ± . The
normalization condition ||ψν± || = 1 allows to fix the modulus of C ± , but not their
phases φ ± = arg C ± . It is clear from the semiclassical formulas that the choice
of φ ± is equivalent to the choice of initial points b± in the integral defining the
related action function on 3± . We explain that the choice of points b± is “true”
from the point of view of formula (1.5) if they are so-called Stokes points.? The
exact constructions and definitions are contained in Section 4.
1.2. THE STRUCTURE OF THE PAPER
The paper is organized as follows. In Section 2 we recall well-known formulas

describing the splitting of eigenvalues for the “double-well” problem and give
their obvious geometric interpretation. In Section 3 we deal with the “momen-
tum” tunneling in the 1-D periodic Sturm–Liouville problem, and we use geomet-
ric language to interpret the corresponding formulas for the eigenvalue splitting.
? Note the “momentum symmetry” in the example considered is different from the “kinematic
symmetry” of the potential in double well problems, and the method of the construction “modes via
quasimodes” based on “kinematic symmetry” (see, e.g., [16, 8]) cannot be used in our case.
We explain also the connection between “momentum" tunneling and band theory
and “over barrier” reflection. These two sections include well-known 1-D results,
which we only interpret from the point of view of their generalization to the mul-
tidimensional problem. Section 4 contains the main results of the paper and is
devoted to the Laplace–Beltrami operator on a torus; we give the exact definitions
and descriptions of geometrical objects for the splitting of eigenvalues in (1.6) as
well as for the asymptotics (1.5) of true eigenfunctions corresponding to a pair of
Liouville tori.
The preliminary version of this paper appeared in [11]; in [12] the results were
generalized for Laplace–Beltrami operators on 2D-surfaces with quadratically in-
tegrable geodesic flow.
Let us stress again that it was not difficult to obtain the analytical expressions
for the values and functions in formulas (1.5) and (1.6) by means of results [13] and
[18], but it required some calculations to see them and to catch their geometrical
structure. The main difficulties were strict and suitable definitions and descriptions
of “the oriented tunnel path” (this defines the phase ρ(3+ , 3− ) in (1.6)) and the
construction of the projection of complex singularities of the Lagrangian manifold
to (real) invariant tori (this defines the vector m in the amplitude in (1.6)). Let us
add that the structure of formula (1.6) and its interpretation looks natural enough,
and therefore we hope that (1.6) should hold in other concrete problems. Also we
hope that the objects introduced in Sections 2–4 can help to obtain formulas (1.5),
(1.6) and similar formulas in other examples directly from the multidimensional
complex WKB-method.
NOTATIONS. Everywhere in this paper we denote by 3 the Liouville tori of

classical Hamiltonian systems (these tori can be one- or two-dimensional). Asymp-
totic eigenvalues and eigenfunctions (quasimodes) of Schrödinger-type operators
are denoted by Eν , 9ν respectively (maybe with indices “+” or “−”) and “true”
(exact) eigenvalues and eigenfunctions (modes) by E eν , 9
eν .
2. Tunneling in the One-Dimensional Double-Well Problem

Here we recall well-known facts connected with the quantum double-well problem
(see, e.g., [22]). Let Q = R and H (x, p) = p 2 /2 + V (x) where V (x) is an even
polynomial: V (−x) = V (x). Let V (0) be a local maximum of V ; if min V <
E < V (0), then the level set H (x, p) = E consists of (at least) two circles 3±
(Figure 1).
If Eν satisfies the Bohr–Sommerfeld quantization rule
Z Z x+ p
1 1
p dx = 2(E − V (x)) dx
2π h 3+ π h x−

1 1
= ν + , ν ∈ Z, ν = O , (2.1)
2 h
Figure 1. Invariant circles 3± in a classical double-well problem.
then there exist two functions 9ν± (9ν+ (x) = 9ν− (−x)), satisfying up to O(h2 ) the
corresponding Schrödinger equation:

h2 d2
− + V (x) 9ν± = Eν 9ν± + O(h2 ).
2 dx 2
Here x ± are the points of intersection of the circle 3+ with the x-axis. However
it is well known ([22, 13, 14]), that the exact spectrum of the Schrödinger opera-
tor is nondegenerate and the true eigenfunctions 9 eν± correspond to two different
eigenvalues Eeν± ; the difference between these eigenvalues has the form
Z x− p !
eν− − E
eν+ = ωh 1
E exp − 2(V (x) − Eν ) dx (1 + O(h)) (2.2)
π h −x −
eν± up to O(h) coincide with the odd and even linear combina-
and the functions 9
±
tions of 9ν :
eν± = 9ν+ ± 9ν− + O(h).
9
Here ω = dH/dI is the frequency of classical motion along the circle 3+ (or 3− ),
I is action variable.
Now we give an obvious geometric interpretation of the formula (2.2) (see, e.g.,
[6]). Consider the complexification 8 of the real plane R 2 ; we denote by (z, w)
complex coordinates, corresponding to real coordinates (x, p), respectively. In 8
consider the curve (Riemannian surface) L, defined by the equation H (z, w) = Eν .
On this surface, consider a loop 0 (we will call this loop a tunnel cycle) such that
the projection of this loop onto the z-plane coincides with the segment −x − , x − of
the real axis. The surface L and the loop 0 are plotted on Figure 2. The surface is
obtained by gluing together two copies of the complex plane, each cutt along the
segments [−x + , −x − ] and [x − , x + ].
Then the formula (2.2) can be rewritten in the form:
eν+ = hω e−ρ/ h (1 + O(h)),

eν− − E
E
π
Figure 2. Tunnel cycle on the Riemannian surface.
where
Z
1
ρ= Im(w dz).
2 0
We call ρ the tunnel distance between the circles 3± .
3. “Momentum” Tunneling in the One-Dimensional Case

3.1. RELATIONSHIP WITH BAND THEORY OF THE PERIODIC
STURM – LIOUVILLE PROBLEM AND OVER BARIER REFLECTION
All considerations about “inverse” tori from subsection are also true for the Schrö-
dinger-type operator −h2 1 + V (x) on Q, where V (x) is a potential. Let us discuss
here the 1-D case and recall the relationship between the 1-D spectral problem
with tunneling and the band theory of the periodic Sturm–Liouville problem. If
Q = R, the one-dimensional invariant tori (circles) 3+ and 3− coincide and the
mentioned questions do not arise. But if one considers the 1-D periodic problem
for the Schrödinger equation (i.e., the Schrödinger equation on the circle, V (x) is
a periodic function), then the phase space T ∗ Q is a cylinder and if the classical
energy level exceeds max V (x), then 3+ and 3− do not coincide and the question
about splitting makes sense. Generally speaking, the answer to this question is well
known and can be easily understood in terms of the spectral properties of a periodic
Sturm–Liouville operator in L2 (R).
Actually (see, e.g., [17, 23]), the spectrum of the Schrödinger operator H b =
−h (d /dx )+V (x) with periodic coefficient V (x) (the Sturm–Liouville operator)
2 2 2
in L2 (R) is a union of segments ( “allowed” zones) on the real axis, divided by

gaps. At least one of the solutions of the Schrödinger equation at the boundary
point of each gap is a periodic or anti-periodic function; if the gap has a non-
zero length (i.e., is not degenerated to a point) then the second solution is neither
periodic nor anti-periodic. The circles 3+ , 3− on the cylinder correspond to pe-
riodic asymptotic solutions of the Schrödinger equation, and the values Eν belong
to neighborhoods of gaps with even numbers k = O(1/ h). If the gaps are non-
degenerate, true (exact) eigenvalues of the periodic problem are simple and the
distance between two different eigenvalues is equal to the length of the gap. The
length δk of the kth gap tends to zero as k → ∞; if V (x) is an infinitely smooth
function then δk = o(k −N ) for arbitrary N > 0 (see, e.g., [23, 13]). So in this
case (non-degenerate gaps), each asymptotic eigenvalue corresponds to a pair of
exact eigenvalues (each coinciding with one end of the gap); the distance between
these two eigenvalues is o(hN ). This implies the appearance of the “asymptotically
degenerate” spectrum of the operator H b on the circle, which we mentioned above.
The degeneration of gaps is a non-generic phenomenon, although the finite gap
case (only a finite number of gaps are non-degenerate) plays a very important role
in the theory of integration of nonlinear equations (see, e.g., [10, 23, 29]).
In other words, the quasimodes 9ν+ and 9ν− associated with the circles 3+ and
3− , asymptotically describe an invariant subspace of the operator H associated
with two ends of the gap of its spectrum in L2 (R). Of course, all this is true for the
operator H b = 1/c2 (x)d2 /dx 2 on the unit circle, where c2 (x) 6= 0 is a smooth 2π -
periodic function. If the potential V (x) or the function c2 (x) are analytic functions
in some neighborhood of the real axis (for instance, if V (x) is a trigonometric
polynomial), then the formula for splitting or for the length of the gaps is known
(see [13]). This formula describes the tunneling between the circles 3+ and 3−
and includes integrals along certain paths or cycles in the complex phase space;
they take the role of the instanton in this case (see below). Conditions under the
potential when this formula applies are the sufficient ones of non-degeneration of
the corresponding gap. In Appendix 3 (see Assertion A3.2 and the last paragraph)
we describe the algorithm enabling at least numerically to check their validity.
This “splitting” effect is close to the well-known quantum over barrier re-
flection (see, e.g., [22, 5, 26]). We emphasize again that the principle difference
between these examples and 1-D double well problem is that the splitting formula
here really includes the complex paths on a complex plain. Also the supp ψν =
supp ψ̄ν , and π 3+ = π 3− , but the projections of 3+ and 3− onto the momentum
axis do not coincide. That is why we call this type of splitting a “momentum
tunneling”.
3.2. THE FORMULAS AND THE GEOMETRICAL INTERPRETATION
Let Q = S 1 and H (x, p) = p 2 /c2 (x), where c(x) is a positive 2π -periodic

function. For the sake of simplicity we will suppose that c2 is a trigonometric
polynomial (this assumption is technical and can be replaced by a weaker one). We
will also assume that all the nulls of this polynomial are simple. The phase space
T ∗ Q is diffeomorphic to a cylinder. We choose the energy level E > max c√2 (x); in
this case each level set H (x, p) = E consists of two circles 3± : p = ± Ec(x)
(see Figure 3).
The corresponding Schrödinger equation has the form
h2 ϕ 00 + c2 (x)Eϕ = 0, h → 0. (3.1)
Figure 3. Phase space for the periodic Sturm–Liouville problem.
For each Eν satisfying the Bohr–Sommerfeld quantization rule

√ Z 2π
E
c(x) dx = ν
2π h 0
there exist two distinct functions 9ν± , satisfying the condition

d2
h2 2 + c2 (x)Eν 9ν± = O(h2 ).
dx
Each of these functions corresponds to a certain circle 3± . However, for generic
potentials c2 (x) the spectrum of the problem (3.1) is simple (see, e.g., [17, 23]
and the discussion in Section 1). Asymptotics of the distance between two close
eigenvalues E eν± as well as the asymptotics of the true eigenfunctions 9 eν± are writ-
ten down in [13]. Here, we rewrite the corresponding formulas using geometric
language; proofs of the results of this section are given in Appendix 3 (they follow
from the results of [13]).
Consider the complex phase space 8. This is a quotient space C 2 /2π Z and
therefore is diffeomorphic to S 1 ×R 3 . Just as in the previous section, we will denote
the complex coordinates in 8 by (z, w). In the space 8 consider the complex curve
L – the level set H (z, w) = Eν . On this curve we introduce the following objects.
(1) Holomorphic one-form θ = w dz.
(2) Holomorphic vector field v defined by the Hamilton function H (z, w). Note
that the Hamiltonian field (∂H/∂w, −∂H/∂z) contains singularities on L; in
order to avoid them we choose v in the form (∂q/∂w, −∂q/∂z), where q =
(H (z, w) − Eν )c2 (z) = w 2 − c2 (z)Eν .
(3) Kählerian metric h, i; we choose this in such a way that the field v has a unit
length hv, vi = 1.
(4) Vector field ξ , defined by the metric h, i and the form θ via the following
condition: θ(ζ ) = ihζ, ξ i for arbitrary tangent vector ζ .
Using these objects we define a special cycle on the surface L. Namely, consider
the rest points a of the vector field ξ and separatrices connecting them with the
(real) circles 3± . Evidently, the complex conjugation maps each rest point of ξ
to the rest point. So, on the surface L, there are two involutions: p → −p (pre-
serving the rest points and interchanging 3+ and 3− ) and complex conjugation
Figure 4. Tunnel cycle and anti-Stokes points in the periodic Sturm–Liouville problem.
(preserving the circles and interchanging the rest points of ξ ). Thus each segment
of a separatrix, connecting some rest point a with some point b+ on the circle 3+
induces a cycle, connecting the points a, ā, b+ , b− , b− ∈ 3− (see Figure 4). As a
result we obtain a number of loops on L. Consider integrals of the form Im θ along
each loop (note that the loops carry the natural orientation induced by the field ξ ).
DEFINITION. The loop 0 with the minimal value of this integral is called a
tunnel cycle on L. Points of intersection b± = 3± ∩ 0 as well as their projections
π(b±) to the configuration space are called Stokes points on the circles 3± or on
the circle Q.
Now, we introduce a number (a tunnel distance) between 3+ and 3−
Z
1
ρ= Im θ
2 0
and two functions
Z !
± −1/4 i x
9ν = (Eν c (x))
2
exp θ . (3.2)
h π(b± )
ASSERTION 3.1. Assume that the tunnel cycle on L is unique up to a homo-
eν± and two
topy. Then for every integer ν = O(1/ h) there exist two eigenvalues E
eigenfunctions 9eν of the periodic Sturm–Liouville problem (3.1), satisfying the
±
following conditions:
wEeν± = Eν + O(h),

e − e + ωh ρ
Eν − Eν = exp − (1 + O(h))
π h
2νh2 e−ρ/ h
= R (1 + O(h)), (3.3)
π ( 1 2π c(x) dx)2
2π 0

eν± = 9ν+ ± 9ν− + O(h).
9 (3.4)
Here ω is the frequency of the classical motion along the circle 3+ (or 3− ).
The proof is presented in Appendix 3, and it follows from the formulas of the
book [13].
Remarks. (1) The real and imaginary parts of the Kählerian metric defined
above on the surface L defines there Riemannian and symplectic structures; ac-
cording to these the vector field ξ is simultaneously a locally R Hamiltonian and
locally gradient
R one. The local Hamiltonian for this field is Re θ, and the local
potential is Im θ.
(2) Functions 9ν± are “quasimodes”, corresponding to the circles 3± ; exponents
in (3.2) are defined by the action functions on these circles.
(3) Stokes points b± play an important role in the definition of “quasimodes”
±
9ν . The choice of these points as initial ones for the integrals in (3.2) guaran-
tees that the true eigenfunctions are close to the odd and even (with respect to
permutations) combinations of the quasimodes (see (3.4)). If one chooses initial
points in another way, the true eigenfunctions will be close to some other (more
complicated) linear combinations of the quasimodes.
(4) It is possible to avoid the assumption of uniqueness of the tunnel cycle. How-
ever, if there are many non-homotopic tunnel cycles, the formula for the splitting
must be replaced by another one (see [13]).
(5) The limit h → 0 for eigenvalues of the operator h2 c−2 d2 /dx 2 is of course
equivalent to the limit ν → ∞ for eigenvalues of the operator c−2 d2 /dx 2 (i.e., as
h = 1) where ν is the number of the eigenvalue.
4. Tunneling between Two Nondegenerate Invariant Tori for the

Laplace–Beltrami Operator on a Torus with Liouville Metric
In this and the next two sections we consider the case Q = T 2 – a two-dimensional
torus with Liouville metric
ds 2 = (U1 (x1 ) − U2 (x2 ))(dx12 + dx22 ). (4.1)
As in the previous section, we assume that:
(i1 ) Uj (xj ) are some 2π -trigonometric polynomials and that U1 (x1 )−U2 (x2 ) >
const > 0 for any x1 , x2 (the assumption that Uj are trigonometric polynomials is
made to simplify the statement of the result, it can be replaced by the assumption
of holomorphy in some complex neighborhood of the real axis and periodicity).
The Hamilton function H (x, p) has the form
1
H = (p 2 + p22 ).
U1 (x1 ) − U2 (x2 ) 1
The corresponding Hamiltonian system is integrable: the additional integral is
U2 (x2 )p12 + U1 (x1 )p22
S(p, x) = .
U1 (x1 ) − U2 (x2 )
Invariant (Liouville) tori 3 are defined by equations
H (x, p) = E, S(x, p) = E. (4.2)
In this section we will consider noncritical values of the parameters E, E. The latter
Figure 5. Circles forming two invariant tori of the geodesic flow on a torus.
condition is satisfied for an open set in the plane E, E and means that
dH |3 6k dS|3 .
Equations (4.2) can be rewritten in the equivalent form
p12 + (E − EU1 (x)) = 0, p22 + (EU2 (x2 ) − E) = 0;
each torus 3 is a direct product of two circles S1,2 where Sj on the corresponding
cylinder xj , pj is defined by the equation
pj2 − qj (x) = 0,
q1 (x) = EU1 (x) − E, q2 (x) = E − EU2 (x2 ).
In the phase space T ∗ Q = T 2 × R 2 consider the action of the group Z2 ⊕ Z2 ,

generated by two involutions p1 → −p1 , p2 → −p2 . The image of the torus
3 with respect to these transformations can consist of either two or four tori (the
circle Sj can be mapped by the involution either onto itself, or onto the other circle,
see Figure 5).
In this section, we consider the case of two different tori; the case of four tori is
considered in Appendix 2. In another words, we assume that the parameters E, E
vary in such a way, that max q2 (x) > 0, min q1 (x) < 0 < max q1 (x) or
E
(i2 ) min U1 < < max U1 (4.3)
E
for x ∈ [0, 2π ] (the other case, max q1 (x) > 0, min q2 (x) < 0 < max q2 (x) is
quite analogous).
Now we consider to the corresponding quantum problem. This is the spectral
problem for the Laplace–Beltrami operator −h2 1 on T 2 with metric (4.1); in
coordinates 1 has the form
2
1 ∂ ∂2
1= + .
U1 (x1 ) − U2 (x2 ) ∂x12 ∂x22
Asymptotic (as h → ∞) eigenvalues Eν , ν ∈ Z+2 up to O(h2 ) are defined by

the quantization conditions (see [24, 18])
Z Z
1 1 1
p1 dx1 = ν1 + , p2 dx2 = ν2 , νj = O(1/ h). (4.4)
2π h S1 2 2π h S2
Each pair Eν , Eν , satisfying the latter conditions, define two distinct tori 3± and
therefore corresponds to two different quasimodes 9ν± . These functions can be
expressed in terms of the Maslov canonic operator (see [24, 28]) on the tori 3± .
However, the true eigenvalues of the operator −h2 1 are simple; as in the previous
two sections, we have here two distinct but very close eigenvalues E e± . Our goal
is to calculate the difference (splitting) between these eigenvalues as well as the
asymptotics of the true eigenfunctions 9 e±.
Assume that, in addition to (i1 )–(i2 ) the following condition holds:
(i3 ) Each trigonometric polynomial Uj has only two extreme points (one max-
imum and one minimum) on the real circle and all the nonreal nulls of qj are
simple.
This condition can be replaced by the weaker one with the help of the methods
of [13].
We will describe the result using geometric objects similar to those defined
in the previous section. Namely, consider the complex phase space 8; this is a
quotient space C 4 /(2π Z⊕2π Z) and is diffeomorphic to the direct product T 2 ×R 6 .
In this space, consider a complex Lagrangian manifold L defined by the equations
H (z, w) = Eν , S(z, w) = Eν .
Here (z, w) are complex coordinates in 8 corresponding to (x, p). On this mani-
fold consider the following objects.
(1) A closed holomorphic one-form θ = (w, dz).
(2) A pair of commuting holomorphic vector fields v1 , v2 . We will normalize
these fields by the condition
∂v2 θ(v1 ) = 0.
Note that Hamiltonian fields sgrad H and sgrad S are singular on L; in order to
avoid singularities, one has to use Hamiltonian fields vj = sgrad(wj2 − qj (z)).
(3) A Kählerian structure h, i, defined by these fields; we choose this metric in
such a way that v1 , v2 form a Hermitian orthonormal base in each complex tangent
plane to L: hvj , vj i = 1, hv1 , v2 i = 0.
(4) A vector field ξ , defined by the metric h, i and the form θ:
θ(ζ ) = ihζ, ξ i (4.5)
for each tangent vector ζ .
Now using these objects we define a special cycle on L. Namely, consider the
rest points of the vector field ξ on L and separatrices – 2D surfaces, formed by
Figure 6. Tunnel cycle connecting two tori 3± and anti-Stokes points on the tori.
trajectories coming to the rest points. Consider trajectories of ξ lying on separatri-

ces and intersecting the real tori 3± . Each segment of such a trajectory induces an
oriented loop on L: one has to map the segment using two involutions p → −p
(preserving the rest points of ξ and interchanging the tori 3± ) and complex conju-
gation (preserving the tori and interchanging the rest points). The construction of
loops
R is illustrated by Figure 6. Now for each of these loops consider the integral
Im θ.
DEFINITION. The loop 0 with the minimal value of the integral is called a tunnel
cycle on the complex Lagrangian manifold L. Points of intersection b± = 3± ∩ 0
(as well as their projections to the configuration space) are called Stokes points on
the tori 3± (or on the torus Q).
In order to generalize the formula (3.3) to the two-dimensional case we define
on the real torus 3+ a certain cohomology class and a certain cycle. Namely, con-
sider the Liouville foliation in the neighborhood of 3+ and choose a smooth base
of cycles γ1 , γ2 on the Liouville tori. These cycles define action-angle variables
I, ϕ in the neighborhood of 3+ . Now define the element µ ∈ H 1 (3+ , R) of the
first cohomology space of 3+ by the formula
µ(η) = (ω, m(η)).
Here, ω = ∂H/∂I is a frequency vector corresponding to the base γ and m is a
vector of intersections: mj (η) = γj ◦ η, where ◦ denotes the index of intersection.
It is easy to check that the cocycle µ is independent of the choice of the base γj .
Now let us consider the complex trajectories of the vector fields vj . From the
“real” point of view they are two-dimensional surfaces; a tangent plane to each
surface is generated by the vector fields vj , ivj . We will denote by ξj vector fields
on these surfaces defined by the formula (4.5) where both h, i and θ are restricted
to the surfaces. Consider the separatrices of these vector fields, intersecting the real
tori 3+ (in another words, we consider trajectories of ξj , connecting 3+ with the
set of singularities of L with respect to the projection to the complex configuration
space; the latter set is defined on L by the condition det(∂w/∂z = 0)). The points
of intersection of these trajectories with 3+ form a graph on this torus. On each
R
edge of this graph we can define a function Im θ where the integral is calculated
along the trajectory of one of ξj , connecting the edge with the rest point. This
function is constant along each edge; we select such edges that this function is
equal to ρ/2, where ρ is a tunnel distance between the tori 3+ and 3− :
Z
1
ρ= Im θ.
2 0
Here, 0 is a tunnel cycle on L. After such a selection we obtain a cycle η0 ∈
H1 (3+ , Z).
Now we are ready to formulate the result.
ASSERTION 4.1. Let functions Uj satisfy the conditions (i1 )–(i3 ). Let the num-
bers Eν , Eν , defined by the formula (4.4) satisfy (4.3) and let the tunnel cycle on the
complex manifold L be unique up to a homotopy. Then there exist two eigenvalues
e± and two eigenfunctions 9
E e ± of the Laplace–Beltrami operator with the metric
(4.1) satisfying the following conditions
eν± = Eν + O(h2 );
E

eν − E
− eν =
+ µ(η 0 )h ρ
E exp − (1 + O(h)); (4.6)
π h

eν± = K3b++ (1) ± K3b−− (1) + O(h).
9 (4.7)
±
Here K3b ± is the Maslov canonic operator [24, 28] on the torus 3± with the marked
point b± ; b± are Stokes points on the tori 3± (see Figure 6).
Remarks. (1) Trajectories of the gradient-Hamiltonian vector field ξ lying on

separatrices (or separatrices themselves) are natural generalizations of anti-Stokes
lines to the multidimensional case. Arbitrary trajectories of this field are the analogs
of the canonical paths. Of course, in the same way it is possible to define general-
izations of the Stokes lines. Namely, consider the vector field iξ . This vector field
again
R will be simultaneously a Hamiltonian
R and gradient one: the local Hamiltonian
is Im θ and the local potential is Re θ. Trajectories of this vector field lying on
separatrices are the natural analogs of the Stokes lines in the multidimensional case.
In the paper [6], other geometric objects ((2n−1)-dimensional Stokes surfaces in a
2n-dimensional Lagrangian manifold L) are used for a study of the close problems
connected with tunneling.
±
(2) Functions K3b ± (1) are quasimodes of the Laplace–Beltrami operator corre-
sponding to the tori 3± ; the points b± have for these quasimodes the same role as
for the functions 9ν± of the previous section (see (3.2)–(3.4)).
(3) The cycle η0 defines “the action variable responsible for the splitting”; if
one chooses a base of cycles on 3+ in such a way that γ1 ◦ η0 = 1, γ2 ◦ η0 = 0,
then the factor µ(η0 ) in (4.6) becomes equal to ω1 – the frequency, corresponding
to the action variable I1 .
(4) Asymptotics of the eigenvalues of the operator h2 1 as h → 0 is of course

equivalent to the asymptotics of the eigenvalues of the operator 1 as k → ∞ where
k is the number of the eigenvalue.
(5) The factor µ(η0 ) in (4.6) can be defined by using the torus 3− instead of
+
3 .
(6) It is easy to see that there exist two different (but homotopic) tunnel cycles
on L. Of course (due to the fact that the form w dz is closed on L) one may use any
of them to calculate the asymptotics of eigenvalues and eigenfunctions.
Proof of Assertion 4.1. The spectral problem for the operator h2 1 possesses the
separation of variables and is reduced to two 1-D periodic Sturm–Liouville prob-
lems (see [18] and Appendix 1, formulas (A1.3)–(A1.5)). Obviously, the distance
between E eν± is equal to the distance between corresponding eigenvalues in the one-
dimensional problem (A1.4). The latter distance is given by the formulas (3.3) or
(A3.31). So we have only to check that these formulas agree with the definitions
of the geometrical objects from this section. Consider first the exponents in (4.6)
and (A3.31) and let us describe the tunnel cycle on L in detail. The complex phase
space 8 is a direct product 81 ×82 of two-dimensional complex phase spaces, each
of them contains the Riemannian surface Lj ; the complex Lagrangian manifold L
is equal to L1 × L2 . The holomorphic vector field vj is tangent to the surface
Lj × a and has the form vj = sgrad(wj2 − qj (zj )). The Kählerian metric h, i in
coordinates (z1 , z2 ) has the form |w1 |−2 dz1 dz̄1 + |w2 |−2 dz2 dz̄2 . The vector field
ξ in the same coordinates is equal to (i|w1 |2 w̄1 , i|w2 |2 w̄2 ). The rest points of this
field are a1 × a2 , where aj are the turning points on Lj . Now consider the real
tori 3± . These tori are the direct products S1 × S2± . The circle S1 contains two
turning points a11 , a12 and the circle S2± does not contain turning points but contains
the Stokes point b2 (see Appendix 3). The tunnel cycle 0, defined in Section 4, is a
direct product a11 × 02 , where 02 is the tunnel cycle on the Riemannian surface L2 .
So the exponent in formula (4.6) coincides with the exponent in formula (A3.31).
Now we have to compare the factor µ(η0 ) in (4.6) with the corresponding factor
in (A3.31). In order to do this, consider the cycle η0 on the torus 3+ . The graph,
formed by the points of intersection of the separatrices of ξj with the torus 3+
consists of three circles: S2+ × a11 , S2+ × a12 , and S1 × b2 . The cycle η0 coincides
with the circle S1 × b2 . Now let us choose the basis of cycles on 3+ as follows:
γ2 = S2+ × a11 , γ1 = S1 × b2 = η0 . Then, obviously, µ(η0 ) = ω2 and the formula
(4.6) coincides with (3.3) or (A3.31).
Appendix 1. Separation of Variables in the Spectral Problem for the

Laplace–Beltrami Operator
In this section, we will reformulate “in coordinates” the results of the previous
section. First of all, let us discuss the separation of variables for the corresponding
quantum problem. Consider the spectral problem in L2 (Q)
−h2 19 = E9, 9 ∈ L2 (Q). (A1.1)
The spectrum of the operator 1 is discrete, its asymptotic properties for large
E are studied in detail in the paper [18]. Let us recall some known facts from [18].
The first integral S of the classical Hamiltonian system defines an operator

b U2 (x2 ) ∂2 U1 (x1 ) ∂2
S = −h 2
+ , (A1.2)
U1 (x1 ) − U2 (x2 ) ∂x12 U1 (x1 ) − U2 (x2 ) ∂x22
commuting with 1. The last fact enables to separate the variables in the problem
(A1.1) and reduce it to two problems on the circles x1 ∈ [0, 2π ], x2 ∈ [0, 2π ] for
the functions ψ1 (x) and ψ2 (x), x ∈ [0, 2π ]:
h2 (ψ1 )00 + (EU1 (x) − E)ψ1 = 0, (A1.3)
h2 (ψ2 )00 + (E − EU2 (x))ψ2 = 0. (A1.4)
Here E is the same as before and E is an eigenvalue of the operator b S (the

constant of separation of variables). The eigenfunctions 9 of the operator 1 are
related to ψ1 (x) and ψ2 (x) by the formula:
9 = ψ1 (x1 )ψ2 (x2 ). (A1.5)
It is clear, that due to the separation of the variables, both the classical motion
and the related spectral problem are decomposed to two one-dimensional problems
on the circles (xj ∈ [0, 2π )). So to write the formulas, one has only to look over
the different cases and combine the related formulas from the one-dimensional case
(see Appendix 3).
Consider first the corresponding one-dimensional classical mechanics. This is
induced by the Hamilton functions Hj (xj , pj ) = pj2 − qj (xj ), where
q1 = EU1 − E, q2 = E − EU2 . (A1.6)
The corresponding phase spaces (cylinders) T ∗ Qj are foliated by closed trajecto-

ries of the corresponding Hamiltonian vector fields; we will denote by Sj trajecto-
ries, corresponding to the null level sets of the functions Hj :
Hj |Sj = 0. (A1.7)
If there are two such trajectories, we will denote them Sj± : the index “+” corre-
sponds to the positive values of momenta pj and the index “–” – to the negative
ones. Let us denote c = E/E, c1 = max U1 , c2 = min U1 , c3 = max U2 , c4 =
min U2 . The structure of the common level set
H = E, S=E (A1.8)
depends essentially on relations between the constants c and cj (in the previous
section we considered the case c2 < c < c1 ).
Let us introduce the action variables I1 and I2 corresponding to the separation

of variables by the formulas
I
1 p
Ij = qj (x, E, E) dx.
2π
Here the integral is calculated along Sj . In the areas where c2 < c < c1 , c3 <
c < c2 , c4 < c < c3 , we express the Hamiltonian and the integral S via the action
variables
H = H (I1 , I2 ) ≡ E(I1 , I2 ),
S = S(I1 , I2 ) ≡ E(I1 , I2 ).
Differentiation of H on Ij gives the frequencies of quasi-periodic motion on the

tori
ωj (I1 , I2 ) = ∂H/∂Ij .
Let us also note that if c → c1 then I1 → 0 and if c → c4 then I2 → 0 and

the two-dimensional tori degenerate to circles, but the frequencies ωj have limits.
However, if c → c2 or c → c3 then the map Ij defined by Equation (4.1) becomes
discontinuous. So the functions Ij are continuous (and infinitely differentiable) on
the sets c4 6 c < c3 , c3 < c < c2 , c2 < c 6 c1 , but are discontinuous on the sets
E/E ≡ c = c2 ; E/E ≡ c = c3 .
The passage of the parameter c across the numbers c2 and c3 relates to the passage
of the parameters (E, E) across the discontinuity set of the functions I1 (E, E),
I2 (E, E) and to the passage across the separatrices of the family of the trajectories
of the Hamiltonian system with the Hamiltonian H in T ∗ Q. Both the classical
motion and the structure of eigenfunctions in the neighborhood of this set becomes
more complicated (see [18]) and we do not consider this case here.
Now we describe the asymptotics of eigenvalues and eigenfunctions of the
Laplace–Beltrami operator for the different values of the parameter c. We will
always assume that conditions (i1 ), (i3 ) of the previous section are fulfilled. In this
section, we give explicit formulas for the case considered in the previous section
leaving the other cases for the next section. So we assume that
(1) c2 < c < c1 . The Equations (A1.8) define in the phase space two tori
3 = S1 × S2± . Their projections to the configuration space Q coincides with the
±
annulus x1− 6 x1 6 x1+ , x2 -arbitrary on the torus; here x1± are the roots of the
trigonometric polynomial q1 (x) (according to the condition (i3 ) there are only two
roots). The quasimodes 9ν± have the form
±
9ν± = K3b ± (1)
as well as the form of the product of two functions

9ν± = ψ1 ψ2± ,
where
b±
ψ1 = KSb11 (1), ψ2± = KS ±2 (1).
2
Here, K is the Maslov canonical operator on the corresponding manifold, the

points b± are defined in the previous section, b1 ∈ S1 is an arbitrary point and
b± = (b1 , b2± ). The explicit formulas in coordinates have the form
Z !
± 1 i x2 p
ψ2 = exp ± q2 (y, Eν ) dy , (A1.9)
(q2 (x2 , Eν ))1/4 h x20
Z !
1 1 x p π
ψ1 = 1/4
cos q1 (y, Eν ) dy + . (A1.10)
(q1 (x1 , Eν )) h x10 4
Here x20 = π(b2±) is the Stokes point on [0, 2π ] (see Sections 3, 4, 7), x10 =
π(b1 ) ∈ [x1− , x1+ ]. The formula (A1.10) is valid inside the interval (x1− , x1+ ); out-
side this interval the function ψ1 = o(hN ) for all N. In the neighborhood of the
points x1± the function ψ1 has the form different from (A1.10); it can be expressed
in terms of Airy functions (see, e.g., [13, 33, 28, 18]). The true eigenfunctions are
close to the sum and difference of 9ν± (see Assertion 4.1 in the previous section).
Appendix 2. Tunneling between Four Tori and between Degenerate Tori

Now we consider the case
(2) c3 < c < c2 . The Equations (A1.8) define in T ∗ Q four tori
p p
3±,± = (p1 = ± q1 (x1 ), p2 = ± q2 (x2 )) ≡ S1± × S2± .
The projection to Q of each of them coincides with the whole Q. The quasimodes
associated with each of these tori are defined by the formulas
±,±
9ν±,± = K3b ±,± (1) = ψ1± (x1 )ψ2± (x2 ). (A2.1)
Here, the functions ψj± have the form (A1.9) (for j = 1 one has to replace in this
formula index “2” by “1”). The points xj0 = π(bj±,± ) are the Stokes points (see
Sections 3, 4, 7) and b±,± ∈ 3±,± are the points, which are projected to (x10 , x20 )
by the natural projection π : T ∗ Q → Q. The formulas for the true eigenvalues
eν±,± and eigenfunctions 9
E eν±,± have the form
eν±,± = Eν + O(h2 ), Eν = H (I1ν , I2ν ),
E
I1ν = hν1 , I2ν = hν2 , νj = O(1/ h),
eν+,+
9 = (9ν+,+ + 9ν−,+ + 9ν+,− + 9ν−,− ) + O(h),
eν+,−
9 = (9ν+,+ + 9ν−,+ − 9ν+,− − 9ν−,− ) + O(h),
eν−,+
9 = (9ν+,+ − 9ν−,+ + 9ν+,− − 9ν−,− ) + O(h),
eν−,−
9 = (9ν+,+ − 9ν−,+ − 9ν+,− + 9ν−,− ) + O(h),
ν
eν+,− − E
E eν+,+ = hω2(I ) exp(−ρ2 / h)(1 + O(h)),
π
ν
eν−,+ − E eν+,+ = hω 1 (I )
E exp(−ρ1 / h)(1 + O(h)),
π
ν
eν−,− − E
E eν−,+ = hω2(I ) exp(−ρ2 / h)(1 + O(h)),
π
eν−,− − E eν+,− = hω 1 (I ν )
E exp(−ρ1 / h)(1 + O(h)). (A2.2)
π
Here, Eν is obtained from the quantization rules (4.4), ρj are the tunnel distances
between the circles Sj on the Riemannian surfaces Lj : wj2 = qj (zj ) (see Section 3
and Appendix 3), ωj are frequencies on Sj .
Let us briefly discuss the geometrical meaning of the formulas (A2.2). In this
case there exist four true eigenvalues of the Laplace–Beltrami operator close to the
number Eν . In order to define the distances between these eigenvalues one has to
consider loops on the complex manifold L connecting each pair of tori. In order
to do this, it is convenient to modify the definition of the tunnel cycle by replacing
there the separatrices of the vector field ξ by those of the vector fields ξj . However,
we consider not all separatrices of these fields but only those passing through the
rest points of the complete field ξ . After such a modification we will obtain tunnel
cycles on L connecting each pair of the tori (the minimum of the corresponding
integral must be calculated with respect to the loops connecting a particular pair).
Now using one torus from the pair we obtain the cycle η0 and the cocycle µ for
each pair. Using these quantities we define tunnel distances ρ between pairs of
the tori. It is easy to see that in this case there are two different numbers ρ1 and
ρ2 (let ρ1 < ρ2 ). The exact eigenvalues form two pairs: the distance between the
pairs vanishes as h → 0 as O(exp(−ρ1 / h)) and the distance between eigenvalues
belonging to the same pair as O(exp(−ρ2 / h)).
Now we turn to the next case
(3) c = c1 . In this case E, E are the critical values of the classical momentum
map. The Equations (A1.8) define in the phase space T ∗ Q two circles (degenerate
tori):
p
3± = (p1 = 0, x1 = x10 , p2 = ± q2 (x2 ), x2 ∈ [0, 2π ]) ≡ a 0 × S2± ,
where a 0 = (p1 = 0, x1 = x10 ), x10 is a unique root of the function q1 on [0, 2π )
and S2± is defined by the formula (A1.7) for j = 2. Projections of the circles
3± on the configuration space coincide with the single geodesic – the meridian
x1 = x10 . It is easy to show that this geodesic is stable in the linear approximation
and according to [4, 27] one can construct the related quasimodes 9ν± . Each of
them again is the product ψ1 ψ2± . The functions ψ2± are defined by the formula
(A1.9) and ψ1 has the form
√
−(x1 −x10 )2 ω0 (x1 − x10 ) ω0
ψ1 = e 2h Hν1 √ . (A2.3)
h
q Here, ν1 = O(1) is a positive integer, Hν1 is ν1 th Hermite polynomial, ω0 =

−q100 (x10 ). From the point of view of the Maslov complex germ theory (see [27])
the functions 9ν± can be presented in the form
b±
9ν± = K30± (1). (A2.4)
b±
Here, K30± is the Maslov complex canonical operator ([27]) on the circle 3± with
the marked point b0± = (a 0 , b2± ), where b2± ∈ S2± are Stokes points (see Sections 3,
4, Appendix 3).
Because of the tunneling effects, the “true” asymptotics of the eigenfunctions
eν± of the operator −h2 1 are the odd and even combinations of the functions 9ν± :
9
√
eν± = 9ν+ ± 9ν− + O( h).
9 (A2.5)
eν± have the form
The corresponding eigenvalues E
eν± = H (I1ν , I2ν ) + O(h2 ),
E
ν ν
E eν+ = hω2 (I1 , I2 ) exp(−ρ/ h)(1 + O(h)),
eν− − E (A2.6)
π
where ρ is the tunnel distance between the circles S1 and S2 in (for now) the 2-D
complex phase space,
I2ν = hν2 , I1ν = h(ν1 + 1/2), ω2 = ∂H/∂I2 ,
νj ∈ Z+ , ν2 = O(1/ h), ν1 = O(1).
Remark. Formula (A2.6) can be obtained as a limit case of formula (4.6) as I
tend to their critical values. However, in the double-well problem (see Section 2)
the situation is quite different: the limit of the formula (2.2) as E tends to min V
exists but leads to the wrong value for the splitting (see [9]).
The cases
(4) c4 < c < c3 and
(5) c = c4 are quite equivalent to the cases (1) and (3), respectively.
Appendix 3. Derivation of the “Momentum” Tunneling Formulas in the

One-Dimensional Case
A3.1. Let us consider the problem
h2 ϕ 00 + q(x, E)ϕ = 0, (A3.1)
ϕ(x + 2π ) = ϕ(x), (A3.2)

where h → 0 and we assume again (to simplify the formulas) that q(x, E) is a
2π -periodic trigonometric polynomial, smoothly depending on the parameter E
(or maybe on several parameters E = (E1 , E2 , . . .)) from a certain set ME . In
important particular situations q is
q = E − V (x), (A3.3)
q = Ec2 (x), c > 0, (A3.4)
and (A3.1), (A3.2) is the spectral problem on a unit circle for the Schrödinger
operator and the operator −h2 c−2 (x) dx d2
2 , respectively. Also, let us assume that the
set q(x, E) > 0 for E ∈ ME on the interval [0, 2π ] is either a point or a segment
[x − , x + ], possibly coinciding with [0, 2π ], and q 6= 0 inside [x − , x + ] (compare
with (i1 )–(i3 )). If the set q(x, E) = 0 consists of the unique point x = x0 , then we
assume that this point is a nondegenerate extremal one:
∂ 2q
ω02 = − (x0 , E) 6= 0. (A3.5)
∂x 2
We have already recalled in Section 3 (see, e.g., [17, 23]) certain properties of
solutions to Equation (A3.1) with the function q (A3.3). Let us also add that in
the “allowed” zones one can choose a basis of solutions constituted by the Floquet
solutions (Bloch functions) (ϕ(x), ϕ(x)), where
ϕ(x + 2π ) = ei2πβ ϕ(x). (A3.6)
The number β (Floquet exponent or quasi momentum) is not unique, but one can
always choose ϕ and β in such a way that the integral part [β] coincides with the
minimal number of zeros (oscillations) on any segment [x, x +2π ] of real solutions
associated with the “allowed” zone. We will suppose that β is chosen in this way.
At one of the ends of each allowed zone, (A3.1) possesses a 2π -periodic solution,
and we can number the zones by k, k = 0, 1, 2, . . . , according to the increase of
eigenvalues Ek of the periodic problem. The periodic solutions are associated with
the left ends of the even allowed zones and with the right ends of the odd ones.
All of these facts are true for any 1-D Schrödinger operator on the circle; but our
operator also includes a small parameter h, and further we study properties of the
spectrum and solutions in the limit h → 0.
1 R 2π √max V − V (x) dx.

Figure 7. The function E(I ); Is = 2π 0
Equation (A3.1) corresponds to a one-dimensional classical dynamical system?

with Hamiltonian H = p 2 − q(x, E) in the real phase space T ∗ S 1 , which is the
cylinder with coordinates x ∈ S 1 , p ∈ R, and we are interested in the solutions of
this system which belong to the zero level set of H. It is convenient to introduce,
in addition to the parameter E, the action variable
I
1
I = I (E) = p dx, (A3.7)
2π
√ H
where p = q(x, E), denotes the integral from 0 to 2π , if q > 0 on the whole
circle and the double integral between zeros x ± of q (i.e., over the set q > 0) if q
in some points is negative.
If q(x, E) has the form (A3.4), then
Z 2π !
√ 1
I= E c(x) dx ,
2π 0
R 2π
and it is possible to define the inverse function E(I ) = I 2 /( 2π
1 2
0 c(x) dx) .
If q has the form (A3.3), then the map I = I (E), defined for any E >
min V (x), is single valued, but is neither monotone nor continuous. It has a discon-
tinuity of the first kind at the point E = max V (x). Thus, the ray E > min V (x)
is separated into two parts: min V 6 E < max V and E > max V . The function
E = E(I ) = strictly increases on each of these parts (see Figure 7).
The inverse map is not single-valued and during the quantization of I this leads
to different interpretations of the quantum numbers. Nevertheless, solving Equation
(A3.7) on corresponding parts with respect to E, one obtains the classical energy
E = p 2 + V (x) in terms of the action variable
E = H (I ). (A3.8)
The derivative of H with respect to I is the frequency of the classical motion
ω = ∂H/∂I. (A3.9)
? Equation (3.1) can be written in the form H(−ih d , x)ϕ = 0.
dx
Let us also note that if E → V (x0 ) = min V (x) and x − , x + → x0 (x0 is the
minimum point of V (x)) then the classical system passes to a stationary state (at
the equilibrium point x0 ). However the function (A3.9) has a limit and we have
r r
∂ 2q ∂ 2V
ω0 = − 2 (x0 ) = (x0 ) 6= 0. (A3.10)
∂x ∂x 2
This is not true in the case (A3.4), when “stationary” state means that I = 0, the
motion of all the points from the circle S 1 stop and ω → 0. Our future consid-
erations are local with respect to the parameter E and it is sufficient to assume
that q(x, E) is defined in a certain neighborhood of the fixed point E 0 where there
exists a diffeomorphism I = I (E) to its image and 0 < |∂I /∂E| < ∞.
A3.2. Now, we describe the asymptotic solutions of Equation (A3.1) as h → 0.
The corresponding formulas depend on the structure of the set
H(p, x, E) ≡ p 2 − q(x, E) = 0.
The following situations are possible.
(1) q > 0 and q = 0 at the unique point x0 . We have the “zero-dimensional
invariant torus” (equilibrium point) 30 = (p = 0, x = x0 ). The asymptotics of the
periodic solutions and the spectral parameters E are easily obtained with the help
of the harmonic oscillator approximation. It has the form
√
−(x−x0 )2 ω0 (x − x0 ) ω0
ϕ = ϕν ≡ Cν e 2h Hν √ + O(h ) ,
1/2
(A3.11)
h

1
Eν = H (Iν ) + O(h ) ≡ H (x0 , 0) + hω0
2
+ ν + O(h2 ). (A3.12)
2
Here, ω0 is defined by (A3.10), Hν are Hermite polynomials, Cν are normalization
constants, and ν = 0, 1, . . . are the numbers of the functions ϕν . These are arbitrary
positive integers, but they must not depend on h (this means that ν = O(1) as
h → 0). The formula (A3.11) defines the solution ϕ in an arbitrary small (but
independent of h) neighborhood Uδ of the point x0 . Outside this neighborhood let
us continue this function first to the interval [x0 − π, x0 + π ], multiplying it by
a cut off function. We obtain a finite function ϕ̃ν from L2 (R), and of course it
is not periodic. Now let us continue it periodically to the whole real axis and we
construct a periodic solution belonging to the νth allowed zone. It is clear that we
can do this in another way; namely, continue the function ϕν to the whole axis by
the rule (A3.6) and again obtain an asymptotic solution of Equation (A3.1). This
is not surprising: for instance, in the case (A3.3) this means only that the lengths
of the allowed zones “under the barrier” max V (x) are exponentially small with
respect to the parameter h as h → 0. The length δν of the νth allowed zone is
defined by tunneling effects, and
Z x0 +2π p
h ln δν = − |q| dx + o(1) (A3.13)
x0
(see, e.g., [13]). The lengths of the gaps between these exponentially small zones
are equal to ω0 h + O(h2). We do not need any deeper analysis of the formulas for
ϕν and δν in this case.
So in this case formulas (A3.11), (A3.12) and (A3.6) allow us to construct the
asymptotics of the problem (A3.1) and (A3.2). The numbers ν come from the
quantization rule for Eν and they are also the numbers of allowed zones related to
these solutions.? Let us also note that it is natural to choose the quasi-momentum
β in the νth zone in the interval β ∈ (ν, ν + 1).
(2) The function q is strictly positive inside the interval
(x − , x + ) ⊂ [x̃ − π, x̃ + π ], x̃ = (x − + x + )/2,
and it is strictly negative in the complement of [x − , x + ] in [x̃ −π, x̃ +π ], and q = 0

only at the ends of the segment [x − , x + ]. This case is related to closed curves in
the phase space T ∗ S (diffeomorphic to circles):
3 = (p 2 − q = 0, x ∈ [x − , x + ]). (A3.14)
The corresponding eigenvalues Eν of the periodic problem can be defined by ob-

taining Iν from the Bohr–Sommerfeld rule

1
Iν = h ν + + O(h2 ), (A3.15)
2
where ν is a natural number, ν ∼ 1/ h, and then calculating E from (A3.7). If q
has the form (A3.3), then we obtain the energy in the form
Eν = H (I ν ) + O(h2 ). (A3.16)
The related eigenfunctions inside the interval (x1 , x2 ) have the form??
Z !
1 1 xp π
ϕν = cos q(ξ, Eν ) dξ + + O(h). (A3.17)
(q(x, Eν ))1/4 h x+ 4
This representation is invalid in the neighborhoods of focal or turning points x −

and x + and one must use there another representation for ϕν , e.g., via Airy func-
tions (see, e.g., [13, 28, 33]). We do not need these formulas and we do not write
them down here. Now we proceed in the same way as in the previous case: we
multiply the function (A3.17) by the cut-off function and then continue the result
by means of the rule (A3.6). Again we obtain the Floquet (Bloch) solutions and
? Actually, because of the cut-off function the constructed functions with number ν have more
than ν zeros in the interval [x0 − π, x0 + π] – they are identically equal to zero outside Uδ . However,
the exact solutions have no zeros in [x0 − π, x0 + π] \ Uδ although they are exponentially small as
h → +0 outside of the neighborhood Uδ of the point x0 . Thus one must count the zeros only inside
Uδ , and there the number of zeros is equal to ν.
?? Note that the phase is the action function on 3.
particularly the periodic solution in the νth zone. The length 1ν of this zone is
equal to (compare with (A3.13))
Z x+ p
h ln 1ν = − q(x, Eν ) dx + o(1) (A3.18)
x−
(see [13]). The lengths of the gaps between the zones with the numbers ν and ν + 1
are equal to the distances between Eν and Eν+1 . By means of the formula (A3.16)
one obtains the lengths of the gaps in the form hω(I ν ) + O(h2 ). Thus again the
Bohr–Sommerfeld quantization rules and formulas (A3.17) allow asymptotics of
periodic solutions of Equation (A3.1) to be constructed and, as before, the number
ν of the solution coincides with the number ν from the quantization rule and with
the number of the allowed zones related to this solution. (Here again one has
to count only those zeros of the solution which belong to the segment [x − , x + ];
compare with the previous case.)
Finally, let us note that formulas (A3.12) and (A3.13) can be rewritten in the
form (A3.16) and (A3.18), respectively. It is possible to show (see, e.g., [37, 18]),
that both of these formulas are true simultaneously for some numbers ν (these
numbers must be “not too big” for the asymptotics (A3.11) and “not too small”
for the asymptotics (A3.17)). Thus the formulas (A3.15), (A3.16) and (A3.18)
give the quantitative characteristics for the “zone structure” in the “under-barrier”
area. In the case (A3.4), the action variable, according to the quantization rule,
“enumerates” the allowed (stable) zones. Of course this is due to the fact that the
zero-level of the function p 2 − q(x, E) possesses only one connected component.
(3) If the parameter E in (A3.4) is close to the maximum of V (x), then formulas
(A3.15)–(A3.18) break down; at least they must be revised. This corresponds to the
fact, that now the classical motion takes place in the neighborhood of the separatrix
defined by the equation
p 2 − q(x, Es ) = 0, Es ≡ max V (x).
Once the energy E passes through the point max V (x), we see a bifurcation of the
phase picture: the curve 3 splits into two smooth curves (circles) 3+ and 3− in
the phase space T ∗ S 1 which are symmetric with respect to the axis p = 0. Let
us note that the separatrix is related to a singular variety (curve) and the standard
WKB-method does not work here. The turning point x + merges with the turning
point x − (on the circle), we have an unstable turning point and the true asymptotics
in its neighborhood is presented by means of the parabolic cylinder functions (see
[30, 18, 33, 37]). We do not consider this case here. But let us also note that from the
point of view of the spectral theory in L2 (R) the passage of the energy E (going
from smaller to bigger) across the separatrix leads to an increase of the allowed
zones and to a decrease of the gaps. Near the separatrix their lengths have the same
order O(h). When the distance E − Es > 0 becomes O(1), the situation becomes
directly opposite to the under-barrier case (E < Es ): the lengths of the “allowed”
zones are of the order O(h) and the lengths of the gaps are O(h∞ ) (see Figure 8).
Figure 8. Gaps (black segments) and “allowed zones” (white segments) on the E-axis.
(4) Now we consider the main case: q(x, E) > 0 for all x. Here, we have two
invariant curves in T ∗ Q symmetric with respect to p = 0:
√
3± = (p = ± q, x = [0, 2π ]). (A3.19)
Obviously, if the motion on 3+ is defined by some functions (P (t), X(t)), then
the motion on 3− is defined by the functions (−P (−t), X(−t)). Let us fix the
orientation on these curves according to the last formulas. Then the action variable
I will be positive on each of them and the energy H in terms of the action variable
has the same form for each family 3± (and, of course, coincide on 3+ and 3−
for each fixed value of E). Now, let us recall the well-known result about the
asymptotics of the solutions to Equation (A3.1) (see, e.g., [13, Ch. III, §9]).
ASSERTION A3.1. (a) For real q(x, E) > 0 and any arbitrary natural N, there
exist two linearly-independent solutions to Equation (A3.1), ϕ + and ϕ − = ϕ̄ + ,
having in each finite interval [a, b] ⊂ R the following asymptotics as h → 0:
ϕ + = ϕN+ + O(hN+1 ), ϕ − = ϕN− + O(hN+1 ) uniformly in x, where
Z xp Z x !
1 i XN
ϕN+ = exp q(x, E) dx + hm αm (x, E) dx
(q(x, E))1/4 h x0 m=1 x0
Z !
1 i xp
= exp q(x, E) dx (1 + O(h)), (A3.20)
(q(x, E))1/4 h x0
ϕN− = ϕ̄N+ .
Here αm are smooth 2π -periodic functions, x0 is an arbitrary real number.
(b) Let I ν be defined by the Bohr–Sommerfeld quantization rule (note that the
Maslov index of the curves 3± is equal to zero)
I ν = hν, ν ∈ Z+ , ν ∼ 1/ h. (A3.21)
ν
Then there exists a set of “corrections” I(k) , k = 2, 3, . . . , such that for any
+ −
N > 0, the functions ϕN and ϕN for
X
N
Eν = E(IÑν ), IÑν = I ν + ν k
I(k) h
k=2
are 2π -periodic functions of x. The function E(I ) is inverse to the function (A3.7).
Let us discuss this result. It is obvious that the functions ϕN+ and ϕN− are associ-
ated with the curves 3+ , 3− respectively. The parameter E changes continuously
(of course the case (A3.3) is included). If E coincides with E(IÑν ), then the func-
tions ϕN+ and ϕN− become 2π -periodic. However, this does not mean that the related
exact solutions ϕ + and ϕ − are also 2π -periodic: the Assertion A3.1 is true for
any segment [a, b], but only if it is finite. In the case (A3.3) this means that the
spectrum (in L2 (R)) of the periodic Sturm–Liouville operator as a rule has the gap
(E eν− ). The ends E
eν+ , E eν± coincide with Eν mod O(hN+1 ). The lengths of the gaps
are O(h ). The functions ϕN+ , ϕN− only define the asymptotics of a certain basis of
∞
2-dimensional invariant subspace of the operator (A3.1), related to its eigenvalues

eν+ and E
E eν− . Let us repeat that because of the tunneling between 3+ and 3− , the
true asymptotics of the real periodic solutions (the eigenfunctions) can be presented
as a certain linear combination of ϕN+ and ϕN− . If q(x, E) has certain analytical
properties (we assume that it is a trigonometric polynomial), then it is possible
(see [13, 33]) to obtain asymptotic formulas for E eν− − E eν+ and true eigenfunctions.
We will write down these formulas a little later.
Before that let us discuss different interpretation of the quantization rules (A3.15)
and (A3.21) from the point of view of zone numbering in the spectrum of the
problem (A3.1) and (A3.3). The number ν in (A3.21) in contrast to (A3.15) si-
multaneously numbers two periodic solutions of Equation (A3.1) and hence two
corresponding allowed zones. Thus the number m of the zone is related to ν by
means of the following equalities:
m = 2ν − 1 (one of the ends of this zone is E eν+ ),
m = 2ν (one of the ends of this zone is E eν− ), and not by equation m = ν,
as in cases (1) and (2). This difference in numeration agrees with the nonsingle-
valuedness of the function E(I ) on I ∈ (min V (x), ∞).
A3.3. The splitting of E eν− is defined by the complex cycles and paths
eν+ and E
between 3+ and 3− on the complex curve. The answer appeals to such objects as
Stokes and anti-Stokes lines (see [13]). But here we intend to interpret this answer
in a way little different from [13, 33] and close to [26]. Let us introduce certain
geometric objects. Let 8 be the complex 2-dimensional phase space, associated
with T ∗ S 1 : z = x +iy, w = p+iv are the coordinates on 8: x ∈ [0, 2π ) ≡ S 1 , y ∈
R, p ∈ R, v ∈ R. So 8 is diffeomorphic to S 1 ×R 3 , and the complex configuration
space is the cylinder S = S 1 × R. Let us consider in 8 a 1-dimensional complex
Lagrangian manifold L (complex curve) defined by the equation
w 2 = q(z, E). (A3.22)
It is clear that the intersection of L and the real phase space T ∗ S 1 consists of 3+
and 3− . Let us recall that the zeros zj of the function q(z, E) are called the turning
points. Obviously if z∗ is a turning point then the complex conjugated number z∗ is
also a turning point. Each turning point z∗ determines so-called Stokes lines, which
consist of analytic curves defined in the complex configuration space (i.e. on the
cylinder) S by the equation
Z zp
∗
Im S(z , z) ≡ Im q(ξ, E) dξ = 0. (A3.23)
z∗
According to the conditions on the function q, the canonical domain D of the band
type (an annulus) in S (see [13]) exists and has the following properties:
(a) The boundary ∂D of the domain D consists of certain Stokes lines and
includes two connected components.
(b) It is symmetric with respect to the real axis x (the circle S 1 ).
(c) There are no turning points and Stokes lines inside D.
(d) Certain turning points, say zj and zj , Im zj > 0 belong to the upper and
lower connected components of the boundary ∂D. These points are the singular
ones of the boundary: ∂D is not smooth at these points (in the generic case only
one pair of the turning points z0 and z0 belong to ∂D).
The closed domain D is covered by two leaves of the manifold L which is
the Riemannian surface defined by Equation (A3.22). These leaves coalesce in D
only at the turning points zj . Let us fix in√D the continuous branch of the function
√
q(z, E)√by means of the assumption q(z, E)|z=x > 0, x ∈ S 1 and denote
W (z) = q(z, E). According to the ideas of the saddle-point method (see, e.g.,
[14]), let us consider in D two vector fields iW (z) and iW −1 (z) = iW /|W |2 (see
also [32]). It is easy to see that the latter field coincides with the projection of the
field ξ , defined in Section 3 with the help of the Kählerian metric h , i. However
the field iW −1 is singular in the turning points while the corresponding field ξ is
smooth everywhere on the complex curve L. Obviously the vector fields iW (z)
and iW −1 (z) are parallel on D and the difference between them is connected with
a choice of the time t on their trajectories. Let gt0 and gt be the related shifts along
their trajectories in the domain D during the time t and let x 0 be a point on the
circle S 1 = [0, 2π ). In the domain D let us consider the complex phase (the action
function)
Z z
S(x , z) =
0
W (z) dz, (A3.24)
x0
and denote
Z e
z
ρ = 2 Im W (z) dz, (A3.25)
x0
where z̃ is an arbitrary point on the upper part of ∂D and the path of integration
belongs to D. Obviously ρ does not depend on the choice of x 0 and z̃.
ASSERTION A3.2. (a) The trajectories of the vector fields iW (z) and iW −1 (z)
in D coincide with the lines of the steepest ascent of the function Im S(z, x 0 ) (with
respect to z):
∂Im S ∂Im S
iW = +i .
∂x ∂y
(b) Re S(x 0 , z) is constant on these trajectories.

(c) All the trajectories Z(t, x 0 ) of the vector field iW −1 (z) starting from the real
circle x ∈ [0, 2π ) do not intersect ∂D during the time interval |t| < ρ/2. They
reach the boundary ∂D (which consists of Stokes lines) at the instant t = ρ/2 (the
upper part) and at the instant t = −ρ/2 (the lower part). In other words, ∂D =
gρ S 1 ∪ g−ρ S 1 . The time t on the trajectory Z(t, x 0 ) coincides with S(x 0 , Z(t, x 0 )).
(d) The trajectory γ : (z = Z(t, x0 ), 0 6 t 6 ρ/2) which reaches the turning
point zk is the anti-Stokes line, i.e.,
Re S(x0 , z) = 0.
Let us call these trajectories γk the anti-Stokes paths and the pre-images x 0 = x 0,k ,
of the points zk and zk on the circle S 1 (the beginnings of the anti-Stokes paths γk )
the Stokes points.
(e) All these statements remain valid if one simultaneously changes the signs of
the vector fields iW and iW −1 and of the time t. √
(g) One can lift the trajectories Z(t, x 0 ) to the “positive” leaf w = q(z, E) of
the complex Lagrangian
√ manifold L and the trajectories −Z(−t, x 0 ) to its “nega-
tive” leaf w = − q(z, E). Then one obtains the trajectories of the vector field ξ
(see Section 3).
(h) The lifts of the anti-Stokes paths on the surface L form the loops 0k con-
necting 3+ and 3− – the separatrices of ξ on L.
Proof. All the statements easily follow from the definitions of the Stokes lines
and of the domain D. We need only to discuss the behavior of the trajectories in the
neighborhood of the turning points. Let the multiplicity of the turning point zk (the
zero of q(z, E)) be equal to n. Then the study of the trajectories in its neighborhood
also becomes trivial if one changes the variable:
z = z(g) ≡ zk + g 2/(n+2),
where one must choose the branch of the function g 2/(n+2) by means of the condi-
tion of continuity of the related trajectory. 2
We call the cycle 0k a tunnel cycle and the number ρ = ρ(3+ , 3− ) a tunnel
distance between 3+ , 3− . It is easy to see that this definition agrees with those of
Sections 2–4.
Remark. It is easy to see that the tunnel distance can be presented in the fol-
lowing equivalent forms:
Z zj p
ρ= q(z, E) dz, (A3.26)
zj
where the integrating path is an arbitrary one from the domain D, and
I
1
ρ= w dz, (A3.27)
2 0
where 0 is an arbitrary cycle on L homotopic to one of the tunnel cycles. Also one
can see that the Stokes point x 0,k is the (unique) solution of the equation
Z x0 p Z zk p
q(x, E) dx + Re q(z, E) dz = 0. (A3.28)
x 0,k x0
Here, x 0 is an arbitrary fixed point on S 1 .

Now let us formulate an additional assumption on q(x, E). Let us assume that
(i03 ) there exists only one tunnel cycle and the turning points z0 , z0 related to
this cycle are nondegenerate: ∂q/∂z(z0 , E) 6= 0. Of course, this means that only
one corresponding anti-Stokes path exists. Let x 0 be the related Stokes point in the
interval [0, 2π ].
ASSERTION A3.3. Let the trigonometric polynomial q(x, E) > 0 for E from a
certain neighborhood of the point Eν = E(I ν ) (see (A3.7)) and let the assumption
(i03 ) be satisfied. Then in the O(h2 )-neighborhood of the number I ν there exist two
values I+ν and I−ν , I+ν < I−ν such that
h
I−ν − I+ν = exp(−ρ/ h)(1 + O(h)), (A3.29)
π
eν± =
and the 2π -periodic solutions to Equation (A3.1) related to the numbers E
Eν (I± ) + O(h ) have the asymptotics
ν 2
ψ ± = (ϕN+ ± ϕN− )(x, Eν ) + O(hN+1 ). (A3.30)

Here ϕN± have the form (A3.20) and one must take there the Stokes point x 0 as the
point x0 . In particular, if q has the form (A3.3), then
E eν+ = hω (Eν ) exp(−ρ/ h)(1 + O(h)),

eν− − E (A3.31)
π
and if q has the form (A3.4), then
eν+ =
eν− − E 2h2 ν
E 1
H exp(−ρ̃/ h)(1 + O(h)). (A3.32)
π( 2π c(x) dx)2
The tilde over ρ in the last formula means that one must only consider the closed
curves with the energy E = 1 and take this E in the integrals (A3.25)–(A3.27).
The formulas mentioned give the splitting for the exponentially close eigenval-
ues of the Schrödinger operator (A3.1) on the circle. Actually, they show that the
“support” (the wave-front) of the “true asymptotics” of the eigenfunctions belongs
to both curves 3± . In other words, there exists a tunnel resonance between these
two curves, just as in the double-well problem. Let us also note that as in the double
well problem (see [22]), the even combination of ϕN+ and ϕN− corresponds to the
smaller energy level and the odd combination corresponds to the larger one.
Proof. It is contained, for instance, in the book [13] (see Ch. III, §3) and here we
only present some of its parts. According to [13] Equation (A3.1) has the solution

−1/4 1
w1 (z, h) = q (z, E) exp − S(z0 , z) (1 + hε(z)),
h
Z zp
S(z0 , z) = q(z, E) dz,
z0
with the following estimate of the correction ε:

|ε(z, h)| 6 const.
This estimate is valid everywhere except, maybe, on certain Stokes lines. At
least it is valid inside the domain D. By means of the solution w1 (z, h), one can
construct another solution with the asymptotics
w2 (z, h) = w1 (z, h),
also valid in the domain D.
The solutions w1 and w2 have the following form on the real axis:
Z !
−1/4
0 0
− iS(z h,x ) i x
w1 (x, h) = q (x, E)e exp − p dx (1 + hε1 ),
h x0
Z !
−1/4 iS(z0 ,x 0 ) i x
w2 (x, h) = q (x, E)e h exp p dx (1 + hε1 ).
h x0
iS(z0 ,x 0 )
Obviously the functions w1 and w2 multiplied by the constants e h and
0 0
− iS(z h,x )
e are those denoted by ϕN , ϕN , and defined by (A3.20).
The functions w1 and w2 form a basis in the space of solutions to Equation
(A3.1) and hence define the monodromy operator (according to the 2π -shift on the
real axis) with the matrix . One can find this matrix, e.g., in [13]; it has the form

ω11 (h, I ) ω12 (h, I )
= ,
ω12 (h, I ) ω11 (h, I )
where
r 2
2π i
ω11 = exp I + O(h) 1 + exp − ρ (1 + O(h))
h h
(A3.33)
2π i ρ
ω12 = −i exp I− (1 + O(h)).
h h

1 1
The monodromy matrix is equivalent to the Jordan block 0 1 at the points I+ν
and I−ν . This means that
r 2ρ
2π I±ν
Re ω11 ≡ cos + O(h) 1 + exp − (1 + O(h)) = 1, (A3.34)
h h
where ρ is defined by any of the equivalent formulas (A3.25)–(A3.27) (in [13]

the representation (A3.26) is used). This immediately implies the equation I±ν =
νh + O(h2 ), which is equivalent to the Bohr–Sommerfeld rule, as well as the
formula (A3.29) for the distance between the closest solutions I+ν and I−ν in the
neighborhood of the point νh. Equalities (A3.31) and (A3.32) are obtained by
means of the formula E = H (I ). Actually (compare with [6])

eν − E
− eν = H (I− ) − H (I+ ) =
+ ∂H
E ν ν
(I ) + O(I− − I+ ) (I−ν − I+ν ),
ν ν ν
∂I −
which according to ∂H/∂I = ω coincides with (A3.31). Let us prove (A3.30).
Equality (A3.33) gives

ρ
Im ω11 = ± exp − (1 + O(h)), (A3.35)
h
where the plus sign relates to I+ν and the minus sign relates to I−ν . From this one
can find the eigenvectors y + and y − of the matrices (I+ν ) and (I−ν ) respectively

+ + 1 − − 1
y =C , y =C . (A3.36)
1 + O(h) −1 + O(h)
This immediately leads to periodic solutions at the points I+ν and I−ν
w+ = C + (w1 + w2 )|I =I+ν

Z ! !
x
+ −iS(z0 ,x 0 )/ h −1/4 1
= 2C e q (x, E) cos p dx + O(h) ,
h x0
w− = C − (w1 − w2 )|I =I−ν

Z ! !
x
1
= − 2iC − e−iS(z0 ,x 0 )/ h
q −1/4 (x, E) sin p dx) + O(h) .
h x0
From these formulas, one obtains (A3.30) by choosing the constants C ± in a

suitable way. Let us note that equalities (A3.34) and (A3.35) are exact and the
division by exponentially small values does not lead to exponentially large errors
anywhere. 2
Remark. Formulas from Section 3 represent the obvious geometric interpreta-

tion of (A3.31). Formulas from Section 4 and Appendix 2 follow from those of
Section 3 (and (A3.31)) after applying separation of variables (see Appendix 1).
In conclusion let us point out the following fact. The form of the dependence
of q on E is not essential, and all the results mentioned are local with respect to
E. In a certain sense, the parameter I plays a more important role, than E. The
main assumption is that the derivative ∂I /∂E is bounded and bounded away from
zero in a certain neighborhood of the point E. Moreover, it is clear that q may

smoothly depend on several parameters and if q > 0 and q satisfies the conditions
of Assertion A3.3 in the set of these parameters, then the formulas (A3.29) and
(A3.30) are still valid, and the related estimates are uniform with respect to these
parameters on any compact subset.
Also, instead of the trigonometric polynomials one can consider other functions
analytical in a certain neighborhood of the real axis. The assumption of uniqueness
of the tunnel cycle (or the pair of related turning points) also can be avoided, but
then formula (A3.29) must be replaced by another one (see [13]).
Finally, let us repeat that in order to obtain the splitting and the true asymptotic
formulas for eigenfunctions one needs only the tunnel distance and the Stokes
point. It is possible to obtain both of these objects by means of the map gt from
Assertion 3.2. The realization of this idea using a computer is the following: one
shifts the circle S 1 along the trajectories of the vector field iW −1 (z) until meeting
the first singular (turning) point. The time of the meeting is half of the tunnel
distance, and the pre-image on S 1 of the first singular point is the anti-Stokes
point. This procedure automatically chooses the suitable “tunnel” turning points
and obviates the study of other procedures.
Acknowledgements
The first author (S. Yu. Dobrokhotov) is grateful to Yacov G. Sinai, who attracted
attention to the discussed problem. We are also grateful to Anne Boutet de Monvel
for the stimulating discussion about this problem, and to Peter Grinevich, Michael
Karasev, Andreas Knauf, Vladimir Nazaikinski, and especially Alexey Sossinsky
and Leonid Pastur for very useful discussions. The first author is much indebted to
Jochen Bruening and the Volkswagen Fund for the opportunity to do part of this
work at the Humbolt University of Berlin. The second author is grateful to Heiner
Zieschang for the opportunity to do part of this work at the Ruhr University in
Bochum. Also, this work was possible thanks to partial financial support from the
Russian Fund for Basic Research (Grant 96-01-00720).
References
1. Agmon, S.: Lectures on Exponential Decay of Solution of Second-Order Elliptic Equations,
Math. Notes 29, Princeton University Press, 1980.
2. Arnold, V. I.: Modes and quasimodes, Funct. Anal. Appl. 6(2) (1972), 12–20.
3. Arnold, V. I.: Mathematical Methods in Classical Mechanics, Nauka, Moscow, 1974; Engl.
transl. Springer-Verlag, New York, 1989.
4. Babich, V. M. and Buldyrev, V. S.: Short-Wavelength Diffraction Theory. Asymptotic Methods,
Nauka, Moscow, 1972; Engl. transl. Springer-Verlag, Berlin, 1991.
5. Baz, A. I., Zeldovich, Ya. B. and Perelomov, A. M.: Scattering, Reactions and Decompositions
in Nonrelativistic Quantum Mechanics, Nauka, Moscow, 1971; Engl. transl. Israel Program for
Scientific Translations, Jerusalem, 1969.
6. Creagh, S. C.: Tunneling in multidimensional systems, J. Phys. A: Math. Gen. 27 (1994), 4969–
4993.
7. Colin de Verdiere, Y.: Spectre conjoint d’operateurs pseudo-differenttiels qui commitent II. Le
cas integrable, Math. Z. 171(1) (1980), 51–73.
8. Dobrokhotov, S. Yu., Kolokoltsov, V. N. and Maslov, V. P.: Quantization of the Bellman
equation exponential asymptotics and tunneling, Adv. Soviet Math. 13 (1992), 1–46.
9. Dobrokhotov, S. Yu. and Kolokoltsov, V. N.: The double-well splitting of the low energy levels
for the Schrödinger operator of discrete 84 -models on tori, J. Math. Phys. 36(3) (1995), 1038–
1053.
10. Dubrovin, B. A., Matveev, V. B. and Novikov, S. P.: Nonlinear equations KdV-type, finite-gap
linear operators and Abelian manifold, Russian Math. Surveys 31(1) (1976), 55–136.
11. Dobrokhotov, S. Yu. and Shafarevich, A. I.: “Momentum” tunneling between tori and the split-
ting of eigenvalues for the Beltrami–Laplace operators on Liouville surfaces, Preprint Institute
for Problems in Mechanics, 1997.
12. Dobrokhotov, S. Yu. and Shafarevich, A. I.: Tunnel splitting of the spectrum for the Beltrami–
Laplace operators on 2D surfaces with quadratically integrable geodesic flow, Funct. Anal.
Appl. (to appear).
13. Fedorjuk, M. V.: Asymptotic Analysis: Linear Ordinary Differential Equations, Nauka,
Moscow, 1993; Engl. transl. Springer-Verlag, Berlin, New York, 1993.
14. Fedorjuk, M. V.: Saddle Point Method, Nauka, Moscow, 1977.
15. Gultzwiller, M. C.: Chaos in Classical and Quantum Mechanics, Springer-Verlag, New York,
1982.
16. Helffer, B. and Sjöstrand, J.: Puit multiples en limit semiclassic, 2, Ann. Inst. Henri Poincare
42 (185), 127–212.
17. Krein, M. G.: Generalizations of A. M. Lyapunov study about linear equations with periodic
coefficients, Dokl. AN SSSR 14 (1993), 1879–1883.
18. Kosygin, D. V., Minasov, A. A. and Sinai, Ya. G.: Statistical properties of the spectra of the
Laplace–Beltrami operators on Liouville surfaces, Russian Math. Surveys 48(4) (1993), 3–130.
19. Kobayashi, S. and Nomizu, K.: Foundations of Differential Geometry, Interscience Publishers,
London, New York, 1963.
20. Katsnelson, M. I., van Schilfgaarde, M., Antropov, V. P. and Harmon, B. N.: Ab-initio in-
stanton molecular dynamics for the description of tunneling phenomena, Phys. Rev. A (1996),
December.
21. Lazutkin, V. F.: KAM Theory and Semiclassical Approximations to Eigenfunctions, Springer-
Verlag, Boston, Berlin, New York, 1993.
22. Landau, L. D. and Lifshits, E. M.: Quantum Mechanics, Addision-Wesley, Reading, MA, 1965.
23. Marchenko, V. A.: Sturm–Liouville Operators and Their Applications, Naukova Dumka, Kiev,
1977; Engl. transl. Birkhäuser, Basel, 1986.
24. Maslov, V. P.: Perturbation Theory and Asymptotic Methods, Moscow State University,
Moscow, 1965; Engl. transl. Dunod, Paris, 1972.
25. Maslov, V. P.: Nonstandard characteristics in asymptotic problems, Proc. Int. Congr. Math. 1
(1983), 139–184.
26. Maslov, V. P.: The scattering problems in quasiclassical approximation, Dokl. AN SSSR 151(2)
(1963), 306–309.
27. Maslov, V. P.: The Complex WKB Methods in Nonlinear Equations, Nauka, Moscow, 1977;
Engl. transl. Birkhäuser-Verlag, Basel, Boston, Berlin, 1994.
28. Maslov, V. P. and Fedorjuk, M. V.: Semiclassical Approximation in Quantum Mechanics,
Nauka, Moscow, 1976; Engl. transl. D. Reidel, Boston, 1981.
29. McKean, H. P. and van Moerbeke, P.: The spectrum of Hill’s equation, Invent. Math. 30 (1975),
217–274.
30. Marz, Ch.: Spectral asymptotics for Hill’s equation near potential maximum, Asymptotic Anal.
5(3) (1992), 221–267.
31. Mishenko, A. S., Sternin, B. Yu. and Shatalov, V. E.: Lagrangian Manifolds and Canonical
Operator Methods, Nauka, Moscow, 1978.
32. Neishtadt, A. I. and Sidorenko, V. V.: The delayed stability loss in Ziegler’s system, Preprint,
Keldysh Institute of Applied Mathematics RAS, Moscow, 1995.
33. Olwer, F. W. J.: Introduction to Asymptotics and Special Functions, Academic Press, New York,
London, 1974.
34. Polyakov, A.: Quark confinement and topology of gauge theories, Nuclear Phys. B 120 (1977),
429.
35. Rajaraman, R.: Solitons and Instantons, North-Holland, Amsterdam, 1984.
36. Simon, B.: Semiclassical analysis of low lying eigenvalues, Ann. Inst. H. Poincaré 38 (1983),
295–307.
37. Slavyanov, S. Yu.: Asymptotic Solutions One-Dimensional Schrödinger Equation, Leningrad
State University, 1991; Engl. transl. Amer. Math. Soc., Providence, 1996.
38. Thomas, L. E. and Wassell, S. R.: Semiclassical approximation for Schrödinger operators on
two-sphere at high energy, J. Math. Phys. 36(10) (1995), 5480–5505.
39. Wilkinson, M.: Tunneling between tori in phase space, Physika D 21 (1986), 341–354.
40. Wilkinson, M. and Hannay, J. H.: Multidimensional tunneling between excited states, Physika
D 27 (1987), 201–212.
179
Nonclassical Thermomechanics of
Granular Materials
PASQUALE GIOVINE
Dipartimento di Meccanica e Materiali, Via Graziella, Località Feo di Vito,
I-89060 Reggio Calabria, Italy
(Received: 10 September 1998; in final form: 18 August 1999)

Abstract. The description of the flow of a granular material with rigid grains requires a combination
of suggestions from both fluid and solid mechanics owing to the fact that the material has an essen-
tially fluid-like behavior, but it can also be heaped and, moreover, its bulk compressibility depends
on the initial voids distribution. Hence, for the study of this medium, a continuum theory is proposed
here that follows mainly from the thermomechanical approach of Dunn and Serrin, but also takes
into account characteristic postulates about flows of granular materials with inelastic granules; in
particular we obtain the following result: the stress tensor, the kinetic energy and the heat flux must
be additively decomposed. Moreover, the balance of angular momentum is given here in a more
general form suggested by the mechanical nature of the ‘interstitial working’ of Dunn and Serrin.
Mathematics Subject Classifications (1991): 73Q05, 73B05, 73B30, 80A99.
Key words: balance equations, constitutive equations, granular materials, thermomechanics.
1. Introduction
The theory of Dunn and Serrin (1985) concerning the thermomechanics of inter-
stitial working was mainly motivated by the incompatibility of the constitutive
equation for Korteweg fluids with Coleman’s and Noll’s (1963) interpretation of
the second law of thermodynamics.
Furthermore, in the Coulomb model for granular materials with incompressible
grains, the ‘equilibrium’ portion of the Cauchy stress tensor T is a special example
of an elastic material of grade 3; precisely because the volume fraction ν of the
rigid grains is proportional to the bulk mass density ρ of the body, one has
T e = (β0 − β1 ρ 2 + β2 gradρ · gradρ + 2β3 ρ1ρ)I −

− 2β4 gradρ ⊗ gradρ, (1.1)
where grad(·) denotes the spatial gradient, 1(·) the Laplacian (i.e., the trace tr(·) of
grad[grad(·)]) and I the identity tensor; moreover, the βi , for i = 0, 1, 2, 3, 4, are
material constants (see Equation (5.1) of Goodman and Cowin (1971) or Equation
(17.5) of Capriz (1989)).
180 PASQUALE GIOVINE
Moreover, Giovine and Oliveri (1995) proved that, in the conservative case,
granular materials with inelastic grains behave like continua with latent microstruc-
ture (see the definition in Capriz (1989)) for which the description of mechanical
actions requires displacement gradients of higher order than the first, without the
need to introduce an extra ‘order parameter’, as in Goodman and Cowin (1972)
and Giovine (1995).
Hence, these particular bodies must be studied using the broader thermody-
namic proposal of Dunn and Serrin, who posit the existence of a rate of supply of
mechanical energy, the interstitial working, to overcome the inconsistencies that
arise between the consequences of the Clausius–Duhem inequality and the consti-
tutive equations; but, unlike them, here we also modify the mechanical principle of
angular momentum balance, while characteristic assumptions of granular materials
are applied in the balance of linear momentum.
In fact, it is well known that classical kinetic energy is inadequate to describe
the micromotions of the grains relative to one another, even if rotations of the
grains themselves are neglected. Both continuum theory, which introduces a ki-
netic energy of the local dilatational motions, and frictional–collisional theory,
which defines a pseudothermal energy associated with deviations of the motion
of individual particles from the local average velocity, suppose the existence of
a new term in the expression of the total energy (see Equation (2.14) of Bedford
and Drumheller (1983) and Equation (2.2) of Johnson and Jackson (1987), respec-
tively). For consistency, the linear momentum must also be altered by the addition
of the Lagrangian derivative of this extra kinetic energy term.
Moreover, the condition of symmetry of the stress tensor is derived, together
with a new requirement that the static part of interstitial work be divergence–
free, by means of the action–reaction principle: this last restriction assures one
that the material does not support an equilibrium flux of heat–like classical fluids,
but unlike Korteweg fluids.
Finally we observe that we can obtain general properties of the model if we
reduce to study the conservative case; in fact this theory represents a new example
of the general Euler–Poincaré theory of Holm, Marsden and Ratiu (1998) applied
to continuous bodies (cf., the incompressible conservative case studied in Section 6
of Giovine and Oliveri (1995) by means of a Hamiltonian variational principle of
local type). Thus it can be used as an example to generalize that theory in order to
include entropic and dissipative processes.
2. Kinematics
Let R be the set of real numbers, E a three-dimensional Euclidean space on R,

V its translation space, Lin the set of second-order tensors defined on V and Sym
(Skw) the collections of symmetric (skew) tensors of Lin. Let B∗ be in E the region
occupied by the continuum body C in a fixed reference placement, where region
NONCLASSICAL THERMOMECHANICS OF GRANULAR MATERIALS 181
means the closure of an open connected set with smooth boundary ∂B∗ . If x∗ is the
place taken by a material element of C in B∗ , the motion of C is a smooth mapping
x = χ(x∗ , τ ) (2.1)
of C onto a time-sequence of regions in E with the time τ varying in some interval
[τ0 , τ1 ]. The deformation gradient, denoted by
∂χ
F := (x∗ , τ ), (2.2)
∂x∗
is therefore non-singular and has positive determinant indicated with
ι := det F > 0. (2.3)
During the motion the velocity and the acceleration are given by
∂χ ∂ 2χ
v := (x∗ , τ ) and a := (x∗ , τ ). (2.4)
∂τ ∂τ 2
Alternatively, velocity and acceleration at a given time τ can be considered as fields
on the region Bτ := χ(B∗ , τ ) occupied by C in the current placement, so
v = χ̇ (x, τ ) and a = χ̈ (x, τ ). (2.5)
To complete the kinematic specification of the model of a granular continuum
with inelastic grains, the choices of an explicit expression of the principle of mass
conservation and of the kinetic energy of the material are necessary.
The first choice is classical, since it is intended to exclude phenomena of diffu-
sion (typical of the mixtures). The mass of any part S of the continuum body C,
which occupies the subregion P∗ of B∗ in the reference placement, is conserved,
so that, if we call ρ the mass density, the balance of mass is
Z
d
ρ = 0, (2.6)
dτ Pτ
where Pτ := χ(P∗ , τ ).
The local form of (2.6) gives the continuity equation, i.e.
ρ̇ + ρdivv = 0, (2.7)
˙ denotes the material time derivative:
where div(·) := tr[grad(·)] and the dot (·)
˙ := ∂(·) + v · grad(·).
(·) (2.8)
∂τ
From (2.6) the material form can also be obtained, which relates the mass density
ρ with the determinant of the displacement gradient F :
ρι = ρ∗ (2.9)
(ρ∗ is the value of ρ in B∗ ).
The choice of a suitable kinetic energy is more delicate. Besides the absence
of diffusion of grains, we also neglect the contribution due to the rotations of the
granules themselves (see Ahmadi (1982) for a generalized theory which includes
this possibility). Whereas, in addition to the usual translational kinetic energy, we
consider that joined to the ‘dilatancy’ as defined by Reynolds (1885), i.e., the
kinetic energy associated to the variations of the volume of interstitial voids (see
Appendix A for an evaluation of this term in a simple fluctuating motion).
Then the density per unit mass of the total kinetic energy of the material turns
out to be
κ = 12 v 2 + 12 γ (ρ)ρ̇ 2 , (2.10)
where the function γ depends on the microgeometry of the material element and
on the kind of admissible microdeformations. Equation (A.7) of Appendix A gives
γ (ρ) = 19 ξ∗2 ρ∗2/3 ρ −8/3 (2.11)
with ξ∗ a constant length. The quantity 13 γ (ρ)ρ̇ 2 is recognized to be what it remains
in our model of the ‘grain temperature’ of granular materials as introduced by
Johnson and Jackson (1987) in a more general context; the grain temperature satis-
fies their ‘pseudothermal energy’ Equation (2.7): we shall see that, in our inelastic
grains theory, this extra equation reduces to an explicit expression for the rate of
dissipation.
By the way, a further generalization of the constitutive expression for the addi-
tional kinetic energy could cause radical changes in the inferences (see the balance
Equation (3.13) of Giovine and Oliveri (1995) for the micromomentum of a gran-
ular material with elastic grains or the Equation (4.6) of ‘granular heat transfer’ of
the extended theory of Capriz and Mullenger (1995)).
Now, it is worth noting that generalized Lagrangians which include kinetic
energy as in expression (2.10) were considered by Holm et al. (1998) in their
stress tensor formulation of Section 7; hence the principal adiabatic nondissipative
aspects of our theory emerge from the more general framework of that Euler–
Poincaré theory. In particular, the dynamic part of our model acquires the following
additional structure, in the conservative case: the Hamilton’s principle formula-
tion; the Kelvin–Noether circulation theorem; the Lie–Poisson Hamiltonian struc-
ture; the characterization of equilibrium solutions as critical points of conserved
quantities (see also Holm et al. (1985)).
3. Dynamic Balances
The characteristic expression of kinetic energy for granular media with rigid grains
is now used to denote, in the classical balance of linear momentum, fluxes through
the boundary of different physical meaning; a balance that has the usual form
Z Z Z
d
ρv = ρf + T n, (3.1)
dτ Pτ Pτ ∂Pτ
with f ∈ V the body force per unit mass and n ∈ V the exterior normal to ∂Pτ ,
but in which the stress tensor T is the sum of a partial stress tensor T p and of a
tensor of inertia flux T i :

T = T p + T i , with T i := −ρ 2 γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 I (3.2)
(see again Johnson and Jackson (1987) for the collisional–translational contri-
bution to the stress tensor in Equation (2.6) or the review article by Hutter and
Rajagopal (1994) in which many granular theories that split the stress tensor in the
Cauchy equation are reviewed); in (3.2) the prime (·)0 denotes differentiation with
respect to ρ.
The tensor T i represents the Lagrangian derivative (times ρ) of the extra kinetic
energy term 12 ργ (ρ)ρ̇ 2 (see Appendix B for the proof) and Equation (3.1) indicates
that any subbody Pτ acts as if it were a body with variable mass for which the influx
of linear momentum per unit area of ∂Pτ is ρ 2 (γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 )n.
Remark. Again we must note that expression (3.2) for the stress tensor is also
deduced in the stress formulation in Section 7 of Holm et al. (1998), if we suppose
that the partial stress depends on a potential energy σ (ρ) per unit mass (see also
Giovine and Oliveri (1995)). The momentum density m of Equation (7.7) of Holm
et al. (1998) is here m = ρv+grad(ρ 2 γ (ρ)ρ̇), while their stress T̂ of Equation (7.8)
is

T̂ = m ⊗ v + ρ 2 γ (ρ)ρ̇(grad v)T + ρ 2 σ 0 (ρ) − 18 γ 0 (ρ)ρ̇ 2 I ;
finally, their Euler–Poincaré Equation (7.6) for continua in the momentum conser-
vation form, ∂m/∂τ = −div T̂ , coincides with our dynamic balance with T p =
−ρ 2 σ 0 (ρ)I , when the external forces f vanishes and the continuity Equation (2.7)
is used.
Before stating the balance of angular momentum, we specify the proposal of

Dunn and Serrin (1985). They suppose the existence of an influx of mechanical
energy through the boundary of Pτ in excess of the usual supply of mechanical
energy due to the working v · T p n of the surface tractions, namely an interstitial
working, of density u · n, attributable in our theory to grain–boundary collisions as
well as to exchange of granules through ∂Pτ ; u ∈ V denotes the interstitial work
flux vector.
We infer the condition of balance of moment of momentum from the principle
of action and reaction, which requires that the power of internal actions must be
identically zero along any rigid virtual motion. We obtain the expression of the
power of internal actions from the theorem of kinetic energy that, in this context,
assumes a special form.
Precisely, from the local appearance of the balance relation (3.1), i.e.,
ρa = ρf + div T , (3.3)
we take the scalar product of both sides of (3.3) by the velocity v and integrate
along regular motions over a subbody Pτ . Using Equations (2.7) and (3.2)1 , the
relation
Z Z Z Z
d
ρκ = ρf · v + (T + T )n · v −
p i
Tp ·L (3.4)
dτ Pτ Pτ ∂Pτ Pτ
follows, where L = grad v is the spatial velocity gradient.

Moreover, the mechanical nature of the interstitial work suggests the insertion
in (3.4) of terms that take account of the extra flux of mechanical energy, so that
the ‘vis viva’ theorem is written as follows:
Z Z Z Z
d
ρκ − T in · v = ρf · v + (T p n · v + u · n) −
dτ Pτ ∂Pτ Pτ ∂Pτ
Z (3.5)
− (T · L + div u).
p
Pτ
The last integral on the right-hand side of Equation (3.5) is the power of internal
forces, while the first two are, respectively, the power of external body actions and
of external surface actions; on the left-hand side there appear, in succession, the
time-rate of change of the total kinetic energy and the flux of kinetic energy through
the boundary.
The power of internal forces has to satisfy the principle of action-reaction, thus
its density −(T p · L + div u) is zero for all rigid body velocity distributions, i.e.,
T p and u must be such that
T p · L + div u = 0, for all velocity fields of the form
(3.6)
v(x, τ ) = c(τ ) + p(τ ) × x,
with c ∈ V translation velocity and p ∈ V angular velocity.
The consequences of this demand become more evident when we take into
consideration the results of Section 5. In particular, we show that the interstitial
work flux u decomposes into the sum of a dynamic part ud = (z ⊗ I ) · L and a
static part us , the first one linear in the trace of L and the second independent of L
itself:
u = us + ud (3.7)
(see Equation (5.13)). Thus, relation (3.6) is modified as follows:
div us + [T p + (div z)I ] · L + (I ⊗ z) · grad L = 0; (3.8)
but, for rigid velocity fields v = c + p × x, it is L = −ep ∈ Skw, where e is the
Ricci permutation tensor, and grad L = 0, then principle (3.8) gives
−ep · T p + div us = 0, (3.9)
for all p ∈ V. For p = 0, we obtain
div us = 0 (3.10)
and hence it follows from the remaining part of (3.9) that the skew part of the partial
stress tensor T p must necessarily be zero. Moreover, as the inertia flux tensor T i
is spherical, we recover the classical condition of symmetry of the Cauchy stress
tensor T .
Nevertheless, it is worth noting that in our theory, unlike those of Dunn and Ser-
rin (1985) and Goodman and Cowin (1972), the classical condition of symmetry of
T is not assumed in advance but it is inferred by the principle of action/reaction as
well as the condition (3.10), which is absent in other theories on granular materials
with rigid grains and which assure us that the flux of the static part of interstitial
working through the boundary of any subbody Pτ of Bτ vanishes.
4. Thermodynamics
We preserve the purely thermal Clausius–Duhem inequality in exactly its classical
form, but we modify the thermomechanical energy balance consistently with the
condition of balance of mechanical energy formulated in (3.5).
During every admissible thermodynamic process, we suppose the existence of
a density of internal energy ε per unit mass, of a rate of heat generation λ per
unit mass due to radiation or other sources, of a heat flux through the boundary of
vector q ∈ V, of a density of entropy η per unit mass and of an absolute positive
temperature θ.
Also, we postulate the validity, for each process and for each subpart Pτ of Bτ ,
of the following generalized balance of energy
Z Z Z
d
ρ(ε + κ) = ρ(λ + f · v) + [T T v + u − q] · n, (4.1)
dτ Pτ Pτ ∂Pτ
in addition to the classical imbalance of entropy

Z Z Z
d −1
ρη > ρθ λ − θ −1 q · n. (4.2)
dτ Pτ Pτ ∂Pτ
With sufficiently regular processes, which also assure the validity of the kinetic
energy theorem (3.5), the balance (4.1) is replaced by
Z Z Z
d
ρε = (ρλ + T · L + div u) −
p
q · n, (4.3)
dτ Pτ Pτ ∂Pτ
which, given (2.9), implies the local generalized Neumann equation
ρ ε̇ = ρλ + T p · L + div u − div q. (4.4)

Using Equation (4.4) and (4.2) and the usual localizing procedure, the reduced
version of the Clausius–Duhem inequality follows
ρ(ψ̇ + θ̇η) 6 T p · L + div u − θ −1 q · g, (4.5)
where ψ := ε − θη is the Helmholtz free energy per unit mass and g := grad θ.
The system of balance equations in local form (2.7), (3.3), (3.6), (4.4) and
(4.5) regulates the motion and the temperature field of a granular material with
incompressible grains; more generally, it applies to nonclassical continua that have
the equilibrium portion of the Cauchy stress tensor T which behaves as an elastic
material of grade 3, in order to assure a constitutive equation of the type (1.1), and
an augmented kinetic energy like (2.10) (or, alternatively, if the additional energy is
a more generic nonnegative function κ̄ of ρ and ρ̇ with the properties κ̄(ρ, 0) = 0
and ∂ 2 κ̄/∂ ρ̇ 2 6= 0).
We perform the thermodynamic consequences of the quoted system of balance
equations for the constitutive structure of the fields ψ, η, T p , q and u within the
procedure outlined by Coleman and Noll (1963), appropriately adapted to this
context. Substantially, they differentiate the analytical character of the relations:
Equations (3.3) and (4.4) describe the ambient actions on the body, actions which
are represented by the densities f and λ, respectively; (3.6) and (4.5) character-
ize the internal constitutive structure of the continuum and have to be identically
satisfied during every admissible thermodynamic process.
5. Derivation of the Constitutive Relations

We are concerned with granular materials for which the flow behavior can be
considered similar to fluid one, except that its response depends on the reference
placement B∗ through the reference volume fraction and temperature distribution
or, as in our case of inelastic grains, through the reference mass density and tem-
perature distribution ρ∗ and θ∗ , respectively, in virtue of the relation (A.1) (for this
relation see Bedford and Drumheller (1983) or the experimental results of Arthur
and Menzies (1972)); moreover, as pointed out in the introduction, the medium has
to act like a material of grade 3, hence the constitutive proposal is a special case of
‘Maxwellian’ fluid (see Section 125 of Truesdell and Noll (1992)).
In fact we take as independent constitutive variables for granular materials with
rigid grains the fields ρ∗ , ρ, d := grad ρ, S := grad (grad ρ), θ∗ , θ, g and L; the
insertion of the symmetric tensor S among the variables appears to be a novelty
for granular materials, but it is assumed for consistency with the results of the
conservative case studied in Giovine and Oliveri (1995) (see Equation (6.15)),
which allows a stress tensor of the form required by (1.1).
The equipresence principle and the objectivity under a frame-of-reference change
requires that each dependent constitutive field is given by a smooth hemitropic
function of all the constitutive variables := {ρ∗ , ρ, d, S, θ∗ , θ, g, L}, i.e.,
ψ = ψ̃(), η = η̃(), T p = T̃ (), q = q̃(), u = ũ() (5.1)
and
ψ̃(Q ) = ψ̃(), η̃(Q ) = η̃(), T̃ (Q ) = QT̃ ()QT ,

(5.2)
q̃(Q ) = Qq̃(), ũ(Q ) = Qũ(),
for all and for all proper orthogonal tensors Q (Q ∈ Orth+ ), respectively; we
suppose that the smoothness of the tilde functions in (5.1) is sufficient for all the
following developments and that Q in (5.2) depends on and Q as follows:
Q = (ρ∗ , ρ, Qd, QSQT , θ∗ , θ, Qg, Q̇QT + QLQT ).
We may now incorporate the functional dependence of the free energy ψ and
of the interstitial work u expressed in (5.1) in the Clausius–Duhem inequality (4.5)
and use the chain rule, the identity
ḋ = grad(ρ̇) − LT d (5.3)
and the conservation of mass (2.7). Consequently, we may assert that the constitu-
tive fields ψ, η, T p , q and u must be such that the entropy imbalance

0 6 T p + ρd ⊗ ψd + ρ(ρψρ + d · ψd )I · L + (uL + ρ 2 I ⊗ ψd ) · t grad L −
− ρψS · Ṡ − ρ(η + ψθ )θ̇ − ρψg · ġ − ρψL · L̇ + uρ∗ · d∗ + uρ · d + (5.4)
+ ud · S + uS · grad S + uθ∗ · g∗ + (uθ − θ −1 q) · g + ug · grad g
holds for every motion x and for every temperature field θ; here subscripts denote
partial differentiation, while the left exponent t (·) indicates minor left transposition,
i.e., (t grad L)ij k = (grad L)j ik in components.
In the relation (5.4) we can specify the values of θ̇ in R, of ġ, d∗ and g∗ in V,
of Ṡ and grad g in Sym, of L̇ in Lin and of grad S and grad L in the subsets of
third-order tensors symmetric in all places (grad S ∈ Sym3 ) and in the last two
(grad L ∈ Sym32 ), respectively, independently of and arbitrarily, hence (5.4) is
linear in those quantities: thus its fulfillment implies that, at each , the relative
terms must all vanish. Hence, we obtain the following restrictions:
η + ψθ = 0, (5.5)
ψS = 0, ψg = 0, ψL = 0, (5.6)
uρ∗ = 0, uS · s = 0, ∀s ∈ Sym3 , uθ∗ = 0, ug · G = 0, ∀G ∈ Sym, (5.7)
(uL + ρ 2 I ⊗ ψd ) · t r = 0, ∀r ∈ Sym32 , (5.8)
while the remaining part of the reduced Clausius–Duhem inequality is

0 6 T p + ρd ⊗ ψd + ρ(ρψρ + d · ψd )I · L + uρ · d + uD · S +
+ (uθ − θ −1 q) · g. (5.9)
The conditions (5.6) and the smoothness of ψ assure us that ψ is independent

of S, g and L; moreover, its hemitropic dependence (5.2)1 upon d implies
ψ = ψ̂(ρ∗ , ρ, δ, θ∗ , θ), (5.10)
where δ := d · d.
Equation (5.5) recovers the classical Gibbs relation and shows for the entropy η
the same independence of S, g and L of ψ. From (5.7)1,3 we have that
u = û(ρ, d, S, θ, g, L) (5.11)
and thus, for (5.8) and (5.10), also ψδ must be independent of ρ∗ and θ∗ , i.e.,
ψ = ω̂(ρ∗ , ρ, θ∗ , θ) + 12 φ̂(ρ, δ, θ). (5.12)
Relations (5.11) and (5.12), other than (5.2)5 , (5.7)2,4 and (5.8) modified in
accordance with the first two, guarantee for the interstitial work u of a granular
material with inelastic grains a constitutive expression similar to that valid for the
materials of Korteweg type described in Section 4 of Dunn and Serrin (1985). It
follows by (5.8), (5.12) and Appendix A and B of Dunn and Serrin (1985) that u
is the sum of two distinct parts:
u = (tr L)z + us , (5.13)
where
z = −ρ 2 φ̂δ (ρ, δ, θ)d and us = (ω1 d + ω2 Sd) × g + ω3 d × Sd, (5.14)
with ωi , i = 1, 2, 3, scalar-valued functions of ρ, δ and θ. In (5.13) it is evident the
mechanical character of the interstitial work u represented from the dynamic part
ud = −ρ 2 φδ (tr L)d.
The decomposition (5.13) permits us to obtain, from the principle of action and
reaction (3.6), the relation (3.10) and the classical condition of symmetry of the
Cauchy stress
div us = 0, T = TT (or, better, T p = (T p )T ), (5.15)
as seen in Section 3; the first condition puts in evidence that the only additional
influx of mechanical energy through the boundary of subbodies of interest for
granular
R material is that due to the dynamic part of interstitial working, that is
∂Pτ (u · n).
d
Now we can use the expression (5.14) of the static part of the flux u in (5.15)1
to derive the following constraint for functions ωi :
ω2ρ − 2ω1δ + ω3θ = 0, for all ρ, δ and θ ∈ R+ . (5.16)
Also, by inserting (5.15)1 into the local energy law (4.4) and the reduced Clausius–
Duhem inequality (5.9), we obtain the following simplified forms:

ρ ε̇ = ρλ + T p · L − div ρ 2 φδ (tr L)d − div q and (5.17)
p
T + ρφδ d ⊗ d + ρ ρ ωρ + 2 φδ − div(ρφδ d) I · L−
1
− θ −1 q · g > 0, (5.18)
when we apply results (5.7), (5.8), (5.12) and (5.13).

Thus, unlike materials of Korteweg type in Dunn and Serrin (1985), in granular
media the divergence of the static part us of the interstitial work u drops completely
out of the local energy and entropy laws (4.4) and (4.5).
Now we are able to analyze the pseudothermal energy Equation (2.7) of Johnson
and Jackson (1987) which also appears in other frictional-collisional theories for
granular materials as the balance of ‘fluctuation’ energy.
When the grains are inelastic, we noticed in Section 2 that the grain temperature
of Johnson and Jackson (1987) for this theory is our additional kinetic energy term
multiplied by 23 , that is 13 γ (ρ)ρ̇ 2. Moreover, if we recognize our inertia flux tensor
T i and the dynamic part ud of u as the collisional-translational tensor (−σc ) and
the flux qpt of the pseudothermal energy, respectively, and substitute in (2.7) of
Johnson and Jackson (1987), we obtain
d 1 2
3
2
ρ γ (ρ)ρ̇ 2
= T i
· L − div ρ φδ (tr L)d −0 (5.19)
dτ 3
or, for (2.7) and (3.2)2 ,

0 = −div ρ 2 φδ (tr L)d ; (5.20)
thus Equation (5.20) provides us simply the expression of the rate of dissipation
0 due to the inelasticity of collisions between particles for our theory. With this
prescription the true thermal energy law (2.4) of Johnson and Jackson (1987) turns
out to be our energy balance (5.17), when heat sources are absent (λ = 0). As
remarked at the end of Section 2, when the grains are compressible, an additional
balance equation (no longer trivial) is needed because the proper grain mass den-
sity is no longer constant (see, e.g., the micromomentum balance (4)3 of Giovine
(1995)).
6. Equilibrium and Linear Theory

In order to extract further informations from the residual entropy inequality (5.18),
we must specify an appropriate equilibrium state for granular materials. We define
an equilibrium process for the body C as one in which the independent variables g
and L all vanish, i.e.,
g=0 and L = 0. (6.1)
The left-hand side of (5.18), denoted by σ , has a minimum with respect to g
and L in equilibrium. It then follows that
σg (ρ∗ , ρ, d, S, θ∗ , θ, 0, 0) = 0 and σL (ρ∗ , ρ, d, S, θ∗ , θ, 0, 0) = 0. (6.2)
These conditions yield the equilibrium heat flux and partial stress values
q e := q̃(ρ∗ , ρ, d, S, θ∗ , θ, 0, 0) = 0, (6.3)
and
T e := T̃ (ρ∗ , ρ, d, S, θ∗ , θ, 0, 0)

= −ρφδ d ⊗ d + ρ 2 (ρ −1 φδ + φδρ )δ + φδ (tr S) + (6.4)

+ 2φδδ (d · Sd) − ωρ − 12 φρ I,
respectively. From a comparison of (6.3) with Equation (4.9) of Dunn and Serrin
(1985), we may observe that, again unlike Korteweg materials, granular media
cannot support heat fluxes in equilibrium: this follows from the new relation (3.10)
which, with the requirement of symmetry of the stress T , expresses the balance of
moment of momentum. Instead, our material can sustain shear stresses, a property
that is uncommon in ordinary fluids, owing to the first term on the right-hand side
of (6.4), term proportional to the diad (d ⊗ d) which accounts for the Ericksen
stress. The term (−ρ 2 ωρ I ), independent of δ, appears similar to the pressure term
in the stress tensor for a compressible fluid, whereas the configuration pressure
term (− 12 ρ 2 φρ I ) is related to the variation of the volume distribution of granules.
Furthermore, the Hessian matrix of σ must be nonnegative in equilibrium, in
the sense that its determinant and all of its principal sub-determinants have to be
nonnegative or, equivalently, that the inequality
e
σgg · (g ⊗ g) + 2σgL
e
· (g ⊗ L) + σLL
e
· (L ⊗ L) > 0 (6.5)
must be verified for each g ∈ V and L ∈ Lin; the exponent e denotes values of
functions in equilibrium.
By substituting in (6.5) the expressions of the gradients σgg , σgL and σLL eval-
uated at (ρ∗ , ρ, d, S, θ∗ , θ, 0, 0), we obtain

−qge · (g ⊗ g) + θ Tge − ρ 2 φδθ I ⊗ d · (L ⊗ g)−
(6.6)
− qLe · (g ⊗ L) + θTLe · (L ⊗ L) > 0,
which is satisfied if and only if
qge · g ⊗ g 6 0 (6.7)
and

−4θ qge · (g ⊗ g) TLe · (L ⊗ L)
t e T 2 (6.8)
> θ (Tg ) − θρ 2 φδθ d ⊗ I − qLe · (g ⊗ L) ,
for all vectors g and all tensors L.
The term on the right-hand side of (6.8) is always nonnegative, which means
that, for (6.7), the viscosity tensor TLe is positive semi-definite at equilibrium.
The large number of restrictions on the constitutive equations given above are
still inadequate to provide uncomplicated expressions for the unknown fields: the
wide variety of possible choices, each corresponding to different classes of mate-

rial, could be explored with the use of general representation theorems; here, for
a comparison with previous models of granular materials, we reduce to consider
a first approximation to each choice by means of a linear theory, in order to ob-
tain additional and more explicit informations on the fields from inequalities (6.7)
and (6.8).
First of all, we observe that the specific entropy η, the equilibrium part of the
partial stress T e and the dynamic part of the interstitial work ud will be specified
if a particular choice for the representations of the free energy ψ is established.
We impose an affine dependence of ψ on the variable δ and require that the new
function is positive semi-definite for all choices of the variables; thus, from (5.12),
we have
ψ = ω̂(ρ∗ , ρ, θ∗ , θ) + 12 ϕ̂(ρ, θ)δ (6.9)
with
ω̂(ρ∗ , ρ, θ∗ , θ) > 0, ϕ̂(ρ, θ) > 0. (6.10)
We then obtain from (5.5), (5.13) and (6.4)
η = −ωθ − 12 ϕθ δ, u = −ρ 2 ϕ(tr L)d + us ,

(6.11)
T e = − ρ 2 ωρ + ρϕ + 12 ρ 2 ϕρ δ + ρ 2 ϕ(tr S) I − ρϕd ⊗ d.
Remark. For the special case in which
ω̂(ρ∗ , ρ, θ∗ , θ) = ω̄(ρ∗ , θ∗ , θ) + ρ −1 β̄0 + ρ β̄1 and

(6.12)
ϕ̂(ρ, θ) = 2ρ −1 β̄,
with β̄0 , β̄1 and β̄ material constants, we have that the equilibrium Cauchy stress
value, which coincides with the partial equilibrium stress value (6.11)3 , can be
further specialized to

T e = β̄0 − β̄1 ρ 2 + β̄δ + 2β̄ρ(tr S) I − 2β̄d ⊗ d; (6.13)
thus, as requested in the introduction, we may immediately recover the Coulomb

stress model (1.1) for granular materials, if we set
β0 = β̄0 , β1 = β̄1 and β2 = β3 = β4 = β̄. (6.14)
Equation (5.14)2 , taken at the same order of approximation in d as (6.11), is

ûs (ρ, d, S, θ, g) = ω̂1 (ρ, θ)d + ω̂2 (ρ, θ)Sd × g + ω̂3 (ρ, θ)d × Sd, (6.15)
with ω2ρ + ω3θ = 0 for (5.16).

To complete the assumptions for the linear theory one needs hypotheses about
the dissipative parts q d and T d of the heat flux vector q and of the partial stress
tensor T p , respectively. We assume that q d and T d , at first sight, take on their
traditional representations
q d := q − q e = −κ̄g and T d := T p − T e = λ̄(tr D)I + 2µ̄D, (6.16)
with D the strain rate tensor so defined D := 12 (L+LT ) ∈ Sym: but instead, unlike
classical ones, the thermal conductivity coefficient κ̄ and the viscosity coefficients
λ̄ and µ̄ are now scalar-valued functions of ρ∗ , ρ, d, S, θ∗ and θ.
By inserting (6.16) in (6.7) and (6.8), we obtain the following restrictions on
the coefficients κ̄, λ̄ and µ̄ of the dissipative functions:
κ̄ > 0, (6.17)
2
4κ̄(g · g) λ̄ + 23 µ̄ (tr D)2 + 2µ̄tr D − 13 (tr D)I
(6.18)
> θρ 4 ϕθ2 (tr D)2(d · g)2 ,
while the semi-definite positiveness of the viscosity tensor TLe gives
λ̄ + 23 µ̄ > 0 and µ̄ > 0. (6.19)
The inequalities (6.17) and (6.19) are familiar ones from the theory of linear
viscosity, within the observation after relations (6.16). Also, for (6.17) and (6.19),
the condition (6.18) is verified if and only if

4κ̄ λ̄ + 23 µ̄ > θρ 4 ϕθ2 δ; (6.20)
this inequality concerns the granular structure and appears only in non-classical
theories: it sets a limit to the dependence of the free energy coefficient ϕ on the
temperature θ.
In closing, we summarize the field equations for the linear theory of a granular
material with inelastic granules. By inserting (2.11), (3.2), (6.9) and (6.16)1 in
Equations (2.7), (3.3) and (5.17), we have
ρ̇ + ρtr D = 0,
ξ∗2 ρ∗
2/3

ρa = ρf + div T p + grad ρ 1/3 tr Ḋ + 13 (tr D)2 , (6.21)
9
ρ ε̇ = ρλ + T · D − div ρ 2 ϕ(tr D)d + div(κ̄g),
p
where

T p = − ρ 2 ωρ + ρϕ + 12 ρ 2 ϕρ δ + ρ 2 ϕ(tr S) + λ̄(tr D) I −
(6.22)
− ρϕd ⊗ d + 2µ̄D,
ε = ω − θωθ + 12 (ϕ − θϕθ )δ (6.23)
and the scalar fields must satisfy the inequalities (6.10), (6.17), (6.19) and (6.20).
Finally, it is interesting to note that the constitutive expression (3.2) for the
Cauchy stress tensor T , with the partial stress T p given by (6.22), can be recog-
nized rather implicity in Goodman and Cowin (1972) and not enough explicitly in
the joined paper of Cowin and Nunziato (1981).
In fact Equation (4.19)1 of Cowin and Nunziato (1981) for the stress tensor is
2α
TCN = − 3 + λ̂(tr D) I − 2 d ⊗ d + 2µ̄D, (6.24)
ρm
with

ρ 2 π̄ 2α
3 = ρ ωρ +
2 1 2
2
ρ ϕρ δ + ρ γ ρ̈ +
2
(tr D) − ρdiv 2 d , (6.25)
ρm ρm
when some minor changes in notation are made and l = 0: the so-called ‘equili-
brated inertia’ k(:= γ /ρm2 ) is the supposed constant, so we can add to the right-
hand side of (6.25) the null term 12 ρ 2 γ 0 ρ̇ 2 . Moreover, by choosing α := 12 ρρm2 ϕ̄(ρ),
ρ 2 π̄
λ̂ := λ̄ + ρm
and substituting (6.25) in (6.24), we now obtain the completely
explicit
TCN = T = T p + T i , (6.26)
with T p expressed by (6.22) and ϕ = ϕ̄(ρ) in it.

The identification (6.26) raises some puzzles in regard to the constitutive choice
made in Cowin and Nunziato (1981), where S does not appear as a constitutive
variable while in (6.22) it is evident that T p depends on S itself.
Appendix A: The Material Model

The material elements of a granular continuum are thought of as envelopes that fill
the body without voids between them (the macroelements of Bedford and Drumhel-
ler (1983)). Each envelope contains some rigid inclusions (the grains or microele-
ments) such as the case of a suspension of rigid particles in a compressible fluid,
always excluding diffusion of the grains through the envelope. The fluid density is
considered to be negligible compared with the proper constant mass density ρm of
the suspended particles, hence the bulk mass density % of the element equals ρm
times the volume fraction ν (ν ∈ [0, 1]) of the grains
% = ρm ν (A.1)
and the total volume change of the granular material with inelastic grains is only
due to the increase or decrease of the volume of interstitial voids: no additional
independent kinematic variable is then necessary (as introduced in Goodman and
Cowin (1972) or Giovine and Oliveri (1995) for elastic granules).
The effects of relative rotations of the macroelements or of grains themselves

are neglected, so that the allowed motions within the element are merely homoge-
neous radial motions of the spherical crust due to the displacements of the grain
relative to the center of mass of the element itself.
In a very simple example (but in spite of this, indispensable to obtain a first
suggestion for the choice of an appropriate expression of the kinetic energy den-
sity), one can think of the macroelement as a spherical envelope of radius ς∗ in the
reference placement B∗ containing m spherical rigid grains of constant radius ϑ∗
that are initially at the same distance ξ∗ from the center of mass. The mass of a
grain is µ = 43 ρm π ϑ∗3 , while that of the material element is mµ = 43 %∗ π ς∗3 (where
the star (·)∗ refers to values in B∗ ).
It is supposed that the macroelement expands or contracts homogeneously and
that the grains do not diffuse, so the current radial distance ξi of the generic grain
is related to the current radius ς of the sphere by
ς
ξi = ξ∗ for i = 1, . . . , m. (A.2)
ς∗
Thus the total kinetic energy per unit mass of the macroelement is written
1 X
m
κM = 12 z2 + 1
2 4 3
µ(ξ̇i )2 , (A.3)
3
π ς i=1
where here the dot denotes the time derivative and z is the velocity of the center of
mass.
On the other hand, conservation of mass gives the relation
1/3
%∗
ς = ς∗ , (A.4)
%
so that, from (A.2) and (A.4), the rate of change of ξi is
ξ̇i = − 13 ξ∗ %∗1/3%−4/3 %̇, (A.5)
while the mass of a grain is now
4
µ= %π ς 3 . (A.6)
3m
By using (A.5) and (A.6) in (A.3), we obtain the following expression for κM :
1 2 2/3 −8/3 2
κM = 12 z2 + ξ % %
18 ∗ ∗
%̇ . (A.7)
Appendix B: The Additional Linear Momentum

The proof given here generalizes the developments in Giovine and Oliveri (1995)
for the fixed rigid domain.
When a variation δx(x, τ ), with x in Bτ and the time τ in the interval [τ0 , τ1 ]
(τ1 > τ0 ) of duration of the process, is added to the motion x, a corresponding
variation of the domain Bτ and of the fields defined on Bτ ×[τ0 , τ1 ] is implied. The
total variation of the mass density is restricted by the conservation of mass (3.4),
i.e.,
δρ + ρdiv δx = 0. (B.1)
Then, for variations δx and δρ which vanish at the extremes of the interval
[τ0 , τ1 ], the following equality holds
Z τ1 Z Z τ1 Z

δ dτ 1
2
ργ (ρ)ρ̇ =
2
dτ ρ 2 γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 n · δx −
τ0 Pτ τ0 ∂Pτ
Z (B.2)

− δx · grad ρ 2 γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 .
Pτ
In fact, by using the transport theorem, adapted to this context, and the property
that the variation δ and the material time derivative (·) ˙ commute (see (3.10)1 of
Giovine and Oliveri (1995)), we have
Z τ1 Z Z τ1 Z

δ dτ 1
2
ργ (ρ)ρ̇ 2
= dτ 1
2
ρδ γ (ρ)ρ̇ 2
τ0 Pτ τ P
Z 0τ1 Z τ
˙ + 1 γ 0 (ρ)ρ̇ 2 δρ
= dτ ρ γ (ρ)ρ̇ δρ 2
τ0 Pτ
Z τ1 Z (B.3)
= dτ ˙
ρ [γ (ρ)ρ̇δρ] −
τ0 P
Z τ1 τZ
˙ ρ̇] − 1 ργ 0 (ρ)ρ̇ 2 δρ.
− dτ ρ [γ (ρ) 2
τ0 Pτ
From the classical transport theorem and from Equation (B.1), it follows:
Z τ1 Z "Z #τ1
δ dτ 1
2
ργ (ρ)ρ̇ 2 = ργ (ρ)ρ̇δρ −
τ0 Pτ Pτ
τ0
Z Z
τ1
− dτ ρ γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 δρ (B.4)
τ Pτ
Z τ1 0 Z

= dτ ρ 2 γ (ρ)ρ̈ + 12 γ 0 (ρ)ρ̇ 2 div δx,
τ0 Pτ
since δρ vanishes for τ = τ0 and τ = τ1 . Now integration by parts and the use of
the divergence theorem proves the equality (B.2).
Acknowledgements
We thank one reviewer for suggestions and observations about the compatibility
of our work with the general Euler–Poincaré theory. This research was supported
by the Italian M.U.R.S.T. through ‘Fondi per la ricerca scientifica 60%’ and by the
Italian C.N.R. through the contract n.94.04665.CT26.
References
Ahmadi, G.: A generalized continuum theory for granular materials, Internat. J. Non-Linear Mech.
17 (1982), 21–33.
Arthur, J. R. F. and Menzies, B. K.: Inherent anisotropy in sand, Geotechnique 22 (1972), 115–129.
Bedford, A. and Drumheller, D. S.: On volume fraction theories for discretized materials, Acta
Mechanica 48 (1983), 173–184.
Capriz, G.: Continua with Microstructure, Springer Tracts Nat. Philos. 35, Springer, New York, 1989.
Capriz, G. and Mullenger, G.: Extended continuum mechanics for the study of granular flows, Rend.
Accad. Lincei, Mat. 6 (1995), 275–284.
Coleman, B. D. and Noll, W.: The thermodynamics of elastic materials with heat conduction and
viscosity, Arch. Rational Mech. Anal. 13 (1963), 167–178.
Cowin, S. C. and Nunziato, J. W.: Waves of dilatancy in a granular material with incompressible
granules, Internat. J. Engrg. Sci. 19 (1981), 993–1008.
Dunn, J. E. and Serrin, J.: On the thermomechanics of interstitial working, Arch. Rational Mech.
Anal. 88 (1985), 95–133.
Giovine, P.: Termodinamica dei Continui Granulari Dilatanti, Atti XII Congresso AIMETA ‘95,
Napoli, Italy, Meccanica dei Fluidi IV, pp. 287–292, 1995.
Giovine, P. and Oliveri, F.: Dynamics and wave propagation in dilatant granular materials, Meccanica
30 (1995), 341–357.
Goodman, M. A. and Cowin, S. C.: Two problems in the gravity flow of granular materials, J. Fluid
Mech. 45 (1971), 321–339.
Goodman, M. A. and Cowin, S. C.: A continuum theory for granular materials, Arch. Rational Mech.
Anal. 44 (1972), 249–266.
Holm, D. D., Marsden, J. E. and Ratiu, T. S.: The Euler–Poincaré equations and semidirect products
with applications to continuum theories, Adv. Math. 137 (1998), 1–81.
Holm, D. D., Marsden, J. E., Ratiu, T. S. and Weinstein, A.: Nonlinear stability of fluid and plasma
equilibria, Phys. Rep. 123 (1985), 1–116.
Hutter, K. and Rajagopal, K. R.: On flows of granular materials, Continuum Mech. Thermodyn. 6
(1994), 81–139.
Jenkins, J. T. and Savage, S. B.: A theory for the rapid flow of identical, smooth, nearly elastic,
spherical particles, J. Fluid Mech. 130 (1983), 187–202.
Johnson, P. C. and Jackson, R.: Frictional–collisional constitutive relations for granular materials,
with applications to plane shearing, J. Fluid Mech. 176 (1987), 67–93.
Reynolds, O.: On the dilatancy of media composed of rigid particles in contact, Philos. Magazine 20
(1885), 469–481.
Truesdell, C. and Noll, W.: The Non-Linear Field Theories of Mechanics, 2nd edn, Springer-Verlag,
Berlin, 1992.
197
Random Operators and Crossed Products
DANIEL H. LENZ
Fachbereich Mathematik, J. W. Goethe Universität, 60054 Frankfurt/Main, Germany.
e-mail: dlenz@math.uni-frankfurt.de
(Received: 18 February 1999; accepted: 19 August 1999)

Abstract. This article is concerned with crossed products and their applications to random operators.
We study the von Neumann algebra of a dynamical system using the underlying Hilbert algebra struc-
ture. This gives a particularly easy way to introduce a trace on this von Neumann algebra. We review
several formulas for this trace, show how it comes as an application of Connes’ noncommutative
integration theory and discuss Shubin’s trace formula. We then restrict ourselves to the case of an
action of a group on a group and include new proofs for some theorems of Bellissard and Testard on
an analogue of the classical Plancherel theorem. We show that the integrated density of states is a
spectral measure in the periodic case, thereby generalizing a result of Kaminker and Xia. Finally, we
discuss duality results and apply a method of Gordon et al. to establish a duality result for crossed
products by Z.
Mathematics Subject Classifications (1991): 47B80, 82B44, 46L60.
Key words: random operators, dynamical systems, density of states.
Introduction
Families of random operators arise in the study of disordered media. More pre-
cisely, one is given a topological space X and a family of operators (Hx )x∈X on
L2 (G). Here, X represents the set of “all manifestations” of a fixed kind of disorder
on the locally compact Abelian group G [3, 4].
The simplest example of a disordered medium is given by the periodic structure
of a crystal. In this case X is the quotient of G by the subgroup of periods. In the
general case X will not be a quotient of G, but there will still be an action α of G
on X. The fact that all points of X stem from the same kind of disorder structure is
taken account of by requiring the action to be ergodic.
Whereas for a fixed x ∈ X the operator Hx may not have a large symmetry
group, the whole family of operators will be G-invariant. This leads to the study of
this family as a new object of interest. This study is best performed in the context
of C ∗ -algebras. In fact, it turns out that the crossed products G ×α C0 (X) provide
a natural framework for these objects [3, 6, 7, 10, 36].
As it is, one is even led to a more general algebraic structure, viz C ∗ -algebras
of groupoids when studying certain quasicrystals modelled by tilings [21, 26, 27].
But this is not considerd here.
198 DANIEL H. LENZ
In a remarkable series of papers, Bellissard and his co-workers introduced a

K-theory based method called “gap labelling” for the study of random operators
[3 – 5, 19]. Using results of Pimsner and Voiculescu [32], they were able to get a
description of the possible gaps in many important cases.
As K-theory is best known in the cases, where either G is discrete or X stems
from an almost periodic function, much of their work was devoted to these cases.
However, there are many important examples of more general random operators
[11, 12, 28, 29].
This is one of the starting points of this article. In fact, the main purpose of
Sections 1 and 2 is to study the framework of general random operators. This is
done by means of Hilbert algebras. Sections 3 and 4 are then devoted to special
results in the field of random operators. More precisely, this article is organized as
follows.
In Section 1 we introduce crossed products, study two important representations
and revise their basic theory. Special attention is paid to the relationship between
symmetry properties of random operators and direct integral decompositions.
In Section 2 we use Hilbert algebras to the study of the von Neumann algebras
and the C ∗ -algebras of the dynamical systems of Section 1. We use them to in-
troduce a trace on these von Neumann algebras. We show that this trace coincides
with the trace introduced by Shubin for almost periodic operators [36] and with the
trace studied by Bellissard and others for discrete G [3 – 5]. Moreover, we discuss
how it is connected with Connes’ noncommutative integration theory [13].
In Section 3 we study the case that X is a group itself. We study the relation
between the two representations introduced in Section 1. We provide proofs for
some theorems first announced in [6] and [7] (cf. [2] as well), whose proofs never
seem to have appeared in print. Moreover, we revise the Bloch theory for periodic
operators from the algebraic point of view. This point of view has the advantage
that the operators in questions are neither required to have pure point spectrum nor
to have a kernel. We show that the integrated density of states is a spectral measure
in this case for purely algebraic reasons. This generalizes a result of Kaminker and
Xia [25] and simplifies their proof.
Finally, in Section 4, we adapt a method developed by Gordon et al. [22] for the
study of the almost Mathieu operator to general crossed products by Z.
1. The C ∗ -algebra G ×α C0 (X)
To every dynamical system (G, α, X) a C ∗ -algebra can be constructed called the

crossed product of G and C0 (X) and denoted by G ×α C0 (X). If X consists of only
one point, then G ×α C0 (X) is nothing but the group C ∗ -algebra C ∗ (G). We will be
concerned with two special representations of G ×α C0 (X). For further details on
general crossed products, we refer to [31, 38], for details on topological dynamics
and crossed products see [39 – 41].
RANDOM OPERATORS AND CROSSED PRODUCTS 199
1.1. BASIC DEFINITIONS
A dynamical system is a triple (G, α, X) consisting of

− a separable, metrizable, locally compact, Abelian group G, whose Haar mea-
sure will be denoted by ds,
− a separable, metrizable space X,
− a continuous action α of G on X,
Moreover we will need
− an α-invariant measure on X with supp m = X
to define the representations considered below. We emphasize that this measure is
not needed to define the crossed product.
The group G is acting on L2 (G) := L2 (G, ds) by
Tt : L2 (G) → L2 (G), Tt ξ(s) := ξ(s − t), s, t ∈ G, ξ ∈ L2 (G)
and on L2 (X) := L2 (X, dm) by
St : L2 (X) → L2 (X), St ξ(x) := ξ(α(−t ) (x)), t ∈ G, x ∈ X, ξ ∈ L2 (X).
Given a topological space Y , we denote by Cc (Y ) (C0 (Y ), Cb (Y ) resp.) the alge-
bra of continuous, complex valued functions with compact support (vanishing at
infinity, being bounded resp.). Let k · k∞ denote the supremum norm on either of
these algebras. The crossed product G ×α C0 (X) is defined in the following way:
Equipped with multiplication, involution and norm defined by
Z
a ∗ b(t, x) := a(s, x)b(t − s, α(−s) (x)) ds,
G
a ∗ (t, x) := ā(−t, α(−t ) (x)),
Z
kak1 := ka(s, ·)k∞ ds,
G
a, b ∈ Cc (G × X), t ∈ G, x ∈ X,
Cc (G × X) becomes a normed ∗-algebra. In general, this algebra is neither com-
plete nor a C ∗ -algebra. It is easy to see that
kak := sup{kρ(a)k : ρ Hilbert space representation of Cc (G × X)}
defines a C ∗ -seminorm on Cc (G × X). In fact k · k is a norm, as can be seen by
using the representations πx , x ∈ X, to be defined below. The crossed product
G ×α C0 (X) of the dynamical system (G, α, X) is the completion of Cc (G × X)
with respect to k · k. It is immediate from these definitions that every representation
of (Cc (G × X), k · k1 ) has a unique continuous extension to a representation of
G ×α C0 (X). We will be concerned with two special representations and their
direct integral decompositions.
Remark 1. The algebra A := (Cc (G × X), ∗, ∗ , h· | ·i) where involution and
convolution are defined as above and h· | ·i is the inner product on the Hilbert space
200 DANIEL H. LENZ
L2 (G × X) can easily be seen to be a Hilbert algebra (cf. [18]), i.e. to fulfil the
following conditions:
(i) ha | bi = hb∗ | a ∗ i a, b ∈ A;
(ii) ha ∗ b | ci = hb | a ∗ ∗ ci a, b, c ∈ A;
(iii) For a ∈ A the mapping b 7→ a ∗ b is continuous;
(iv) {a ∗ b : a, b ∈ A} is dense in L2 (G × X).
In particular, the action of A on itself from the left yields a representation of
A on L2 (G × X), which can be extended to a representation of G ×α C0 (X) on
L2 (G × X). These considerations will be given a more precise form in Section 2
in order to study G ×α L∞ (X). Now we prefer to introduce two representations,
that allow a direct integral decomposition.
1.2. REPRESENTATIONS OF G ×α C0 (X)
Let π : Cc (G × X) → B(L2 (G × X)) be given by

Z
π(a)ξ(t, x) := a(t − s, αt (x))ξ(s, x) ds, ξ ∈ L2 (G × X),
G
and let for x ∈ X the representation πx : Cc (G × X) → B(L2 (G)) be given by

Z
πx (a)ξ(t) := a(t − s, αt (x))ξ(s) ds, ξ ∈ L2 (G),
G
where B(H ) denotes the algebra of bounded operators on the Hilbert space H .
Then it is easy to see that π and πx , x ∈ X, are continuous representations of
Cc (G × X). Their extensionsR to G ×α C0 (X) will also be denoted by π and πx .
⊕
Identifying L2 (G × X) with X L2 (G) dm, we get
Z ⊕
π(A) = πx (A) dm, A ∈ G ×α C0 (X).
X
This is obvious for A ∈ Cc (G × X) and follows for arbitrary A ∈ G ×α C0 (X) by

density. As G is amenable, even Abelian, the representation π is faithful (cf. 7.7 in
[31]). Therefore we have
G ×α C0 (X) ' π(G ×α C0 (X)) = π(Cc (G × X)).
Thus the crossed product is the norm closure of an algebra of certain integral
operators.
We remark that the mapping X 3 x 7→ πx (A) is strongly continuous for
A ∈ G ×α C0 (X), as can be directly calculated for A ∈ Cc (G × X), and then
follows by density arguments for all A ∈ G ×α C0 (X). The representation π has
two interesting and well known symmetry properties, that will be given in the next
proposition.
PROPOSITION 1.2.1. (a) Let DR be the algebra of diagonalisable operators on
⊕
the Hilbert space L2 (G × X) ' X L2 (G) dm, then
π(G ×α C0 (X)) ⊂ D 0 .
(b) π(G ×α C0 (X)) ⊂ {Tt ⊗ S−t : t ∈ G}0 . Moreover for t ∈ G, x ∈ X and
A ∈ G ×α C0 (X) the formula Tt παt (x) (A) Tt∗ = πx (A) holds.
Proof. (a) This is the fact that π(A) permits a direct integral decomposition.
(b) This can be directly calculated for A ∈ Cc (G × X) and then follows for
arbitrary A ∈ G ×α C0 (X) by a density argument. 2
Moreover, the following is valid.
PROPOSITION 1.2.2. If α is minimal, i.e. Gx := {αt (x) : t ∈ G} is dense in X

for every x ∈ X, then
(i) σ (πx (A)) is independent of x ∈ X for selfadjoint A ∈ G ×α C0 (X),
(ii) πx is faithful for every x ∈ X.
If G acts ergodically on X, then there exists for each selfadjoint A ∈ G ×α C0 (X)
a closed set 6A ⊂ R and a measurable set XA ⊂ X with µ(X − XA ) = 0 s.t.
σ (Ax ) = 6A for all x ∈ XA .
Proof. The first statement is proven in [25], the second one in Section 4 of [3]. 2
We will now give a second representation of G ×α C0 (X). For tˆ ∈ G b let

tˆ
π : Cc (G × X) → B(L (X)) be defined by
2
Z
π tˆξ(x) := a(s, x)ξ(α−s (x))(tˆ−s | ) ds, a ∈ Cc (G × X), ξ ∈ L2 (X),
G
b For tˆ ∈ G
where (· | ·) denotes the dual pairing between G and G. b the mapping π tˆ
is then a representation of Cc (G × X), which has a unique continuous extension
to a representation of G ×α C0 (X), again denoted by π tˆ. For A ∈ G ×α C0 (X)
the mapping tˆ 7→ π tˆ(A) is strongly continuous, as can be seen using the same
arguments as in the case of the mapping x 7→ πx (A). Therefore, we can define a
representation
Z ⊕ Z ⊕ !
ˆt
e
π := π dtˆ : G ×α C0 (X) → B L2 (X) dtˆ ,
b
G b
G
b Let the unitaries W and U be

where we denote by dtˆ the Haar measure on G.
defined by
W : L2 (G × X) → L2 (G × X), W ξ(t, x) := ξ(t, αt (x)), ξ ∈ L2 (G × X)
202 DANIEL H. LENZ
U := (FG ⊗ I )W ∗ : L2 (G × X) → L2 (G b × X),
where FG : L2 (G) → L2 (Gb) is the Fourier transform and I the identity. Then we
R
have e ∗
π = U π U , where L2 (Gb × X) is identified with b⊕ L2 (X) dtˆ.
G
Remarks. (1) The crossed product G ×α C0 (X) is just the group C ∗ -algebra
C ∗ (G), if X consists of only one point. In this case we identify Cc (G × X) with
Cc (G) and we get π(ϕ)ξ = ϕ ∗ ξ = Tϕ ξ, e π (ϕ)ξ = M(F (ϕ))ξ, where Tϕ
denotes the operator of convolution with ϕ ∈ Cc (G) and M(ψ) denotes the op-
erator of multiplication with ψ. This implies e π (G ×α C0 (X)) = M(C0 (G)) b and
π(G ×α C0 (X)) = {F −1 Mψ F | ψ ∈ C0 (G)}. b
(2) The direct integral decomposition of e
π relies essentially on the symmetry
π(G ×α C0 (X)) ⊂ {Tt ⊗ S−t : t ∈ G}0 ,
as can be seen in the following way: Using W (Tt ⊗ I ) W ∗ = Tt ⊗ S−t , one gets
immediately U (Tt ⊗ S−t )U ∗ = Mt ⊗ I, where Mt denotes the operator of multipli-
b Therefore, we have U π(A)U ∗ ∈ {Mt ⊗ I | t ∈ G}0 ,
cation with (t | ·) on L2 (G).
and this implies (cf. 5, Ch. 2, II in [18]) that U π(A)U ∗ has a direct integral
decomposition.
2. The von Neumann Algebra G ×α L∞ (X)

In this section we study the von Neumann algebra G ×α L∞ (X). We will be par-
ticularly interested in determining its generators and its commutant as well as
introducing and calculating a trace on it. In a sense, much more general consid-
erations can be found in [23], where arbitrary von Neumann crossed products are
studied by means of Tomita Takesaki theory of left Hilbert algebras (cf. [38]). The
trace on G ×α L∞ (X) allows to introduce for each selfadjoint operator affiliated
to G ×α L∞ (X) a canonical spectral measure. This spectral measure is called the
density of states. We will discuss the so called Shubins’s trace formula, relating
the density of states to the number of eigenvalues of certain restricted operators.
We conclude the section with the discussion of certain formulas for the trace in the
case that m(X) < ∞ holds.
2.1. DEFINITION AND IMPORTANT PROPERTIES OF A ×α L∞ (X)
Following 7.10.1 in [31] we define the von Neumann crossed product.

DEFINITION 2.1.1. G ×α L∞ (X) := π(G ×α C0 (X))00 .
We will study this algebra by means of the already defined Hilbert algebra A =
(Cc (G × X), ∗,∗ , h· | ·i) (cf. Remark 1 in Section 1.1). We need some notation.
DEFINITION 2.1.2. (a) For A ∈ A let La (Ra ) be the unique continuous operator
with
La ξ = a ∗ ξ, (Ra ξ = ξ ∗ a), ξ ∈ Cc (G × X),
R
i.e. La ξ = G a(s, x)ξ(t − s, α−s (x)) ds for ξ ∈ L2 (G × X) and similarly for Ra .
(b) The unique extension of ∗ : Cc (G × X) → Cc (G × X) to a continu-
ous mapping of L2 (G × X) into itself will also be denoted by ∗ , i.e. a ∗ (t, x) =
a(−t, α−t (x)) for a ∈ L2 (G × X).
(c) An a ∈ L2 (G × X) is called left bounded (right bounded) if the mapping
ξ 7→ Rξ a (ξ 7→ Lξ a) can be extended to a continuous operator on L2 (G × X).
This operator will be denoted be La (Ra ).
(d) L(A) := {La : a ∈ A}00 , R(A) := {Ra : a ∈ A}00 .
The connection between these crossed products and Hilbert algebras is simple.
LEMMA 2.1.3. For a ∈ Cc (G × X) the equality W ∗ π(a)W = La holds.
Proof. For ξ ∈ Cc (G × X) a direct computation yields W ∗ π(a)W ξ = La ξ and
the lemma follows, as La and W ∗ π(a)W are bounded and Cc (G × X) is dense in
L2 (G × X). 2
The lemma and the definitions of G ×α L∞ (X) and L(A) directly yield
THEOREM 2.1.4. AdW : G ×α L∞ (X) → L(A), A 7→ W ∗ AW is a spatial
isomorphism of von Neumann algebras.
Those operators which are inverse images of left bounded operators under AdW
will play an important role.
DEFINITION 2.1.5. (a) A function a ∈ L2 (G × X) is called the kernel of the
operator A ∈ G ×α L∞ (X) if a is left bounded and A = W La W ∗ .
(b) Let K := {A ∈ G ×α L∞ (X) | A has a kernel}.
We study K in the next proposition.
PROPOSITION 2.1.6. (a) The operator A has the kernel a ∈ L2 (G × X) iff

Z
Aξ(t, x) = a(t − s, αt (x))ξ(s, x) ds a.e.
G
holds for ξ ∈ L (G × X).

2
(b) The set K is an ideal in G ×α L∞ (X). For A ∈ K with kernel a ∈

L (G × X) and B ∈ G ×α L∞ (X) the kernel of AB is given by W ∗ B(W a) and
2
the kernel of A∗ is given by a ∗ .

(c) For A ∈ G ×α L∞ (X) with kernel a ∈ L2 (G × X) the operator Ax is a
bounded Carleman operator with kernel
ax (t, s) := a(t − s, αt (x)),
(i.e. (Ax )f (t) = hax (t, ·) | f i a.e. t ∈ G) for a.e. x ∈ X.
Proof. (a) The statement with “ξ ∈ L2 (G × X)” replaced by “ξ ∈ Cc (G × X)”
is easy to calculate. Using that the maximal operator given by the integral expres-
sion is closed, we get (a).
204 DANIEL H. LENZ
(b) As K contains the algebra π(Cc (G × X)) by Lemma 2.1.3, it is strongly

dense in G ×α L∞ (X). The remaining statements could be calculated directly but
also follow from Propositions 2 and 3 in 3, Ch. 5, I of [18].
(c) We set ax (t, s) := a(t − s, αt (x)) for a ∈ L2 (G × X). Using the Fubini
fx
theorem it is easy to see that ax is the kernel of a Carleman operator (cf. [42]) A
for almost every x ∈ X. It remains to show A fx = Ax , a.e. x ∈ X. For η ∈ L (X)
2
and ξ ∈ L2 (G) a short calculation yields

fx ξ(t),
η(x)Ax ξ(t) = η(x)A a.e.
As η ∈ L2 (G) was arbitrary, this implies

fx ξ,
Ax ξ = A a.e. x ∈ X.
fx is closed, we
Using a countable, dense subset of ξ ∈ L2 (G) and the fact that A
conclude (c). 2
We can now characterize G ×α L∞ (X) and its commutant.
THEOREM 2.1.7. (a) G ×α L∞ (X) = W L(A)W ∗ = {Tt ⊗I, W (I ⊗Mv )W ∗ : t ∈

G, v ∈ L∞ (X)}00 .
(b) G ×α̃ L∞ (X) = R(A) = {Tt ⊗ I, W ∗ (I ⊗ Mv )W : t ∈ G, v ∈ L∞ (X)}00
with α̃ : G × X → X, α̃t (x) := α−t (x).
(c) (G ×α L∞ (X))0 = W R(A)W ∗ = {Tt ⊗S−t , I ⊗Mv : t ∈ G, v ∈ L∞ (X)}00 .
Proof. (a) The equality G ×α L∞ (X) = W L(A)W ∗ has already been proven
in Theorem 2.1.4. To prove the second equality we set
C := {Tt ⊗ I, W (I ⊗ Mv )W ∗ : t ∈ G, v ∈ L∞ (X)}.
C 00 ⊂ W L(A)W ∗ : By L(A) = R(A)0 (cf. Theorem 1 in 2, Ch. 5, I of [18]), it is

enough to show C ⊂ (W R(A)W ∗ )0 , i.e.
CW Ra W ∗ = W Ra W ∗ C, a ∈ Cc (G × X), C ∈ C.
This can be calculated directly.

W L(A)W ∗ ⊂ C 00 : For u ∈ Cc (G) and v ∈ Cc (X) and ξ ∈ L2 (G × X) an easy
calculation yields
Z
∗
π(u ⊗ v)ξ(t, x) = (W (I ⊗ Mv )W ξ(t, x)) · u(s)(Ts ⊗ I )ξ(t, x) ds,
G
and this implies π(u ⊗ v) ⊂ C 00 . The desired inclusion follows.

(b) Defining πα̃ by simply replacing α by α̃ in the definition of π , we get
Ra ξ = πα̃ (W a)ξ, a ∈ Cc (G × X), ξ ∈ L2 (G × X).
This implies R(A) = {πα̃ (W a) : a ∈ Cc (G × X)}00 = {πα̃ (a) : a ∈ Cc (G × X)}00 .

As G ×α̃ L∞ (X) = {πα̃ (a) : a ∈ Cc (G × X)}00 by definition of the crossed

product the first equality is proven. The second equality follows by replacing α by
α̃ in (a), i.e. by replacing W by W ∗ .
(c) The first equality follows from (a) and R(A) = L(A)0 . The second equality
follows by (b) and W (Tt ⊗ I )W ∗ = Tt ⊗ S−t . 2
Theorem 2.1.7 yields G ×α L∞ (X) ⊂ D 0 . In particular (cf. 5, Ch.

R ⊕ 2, II in [18]),
∞
every A ∈ G ×α L (X) can be written as a direct integral A = X Ax dm, whose
fibres are uniquely determined up to a set of measure zero. Similarly it can be seen
that for A ∈ G ×α L∞ (X) the equation
Z ⊕
U AU ∗ = Atˆdtˆ
b
G
holds, where the Atˆ are uniquely determined up to a set of measure zero. From now
on we will identify G ×α C0 (X) with π(G ×α C0 (X)). For A in G ×α C0 (X) we
b := U AU ∗ . For A ∈ G ×α L∞ (X) we define
set Ax := πx (A), Atˆ := π tˆ(A) and A
ˆt
the Ax and A by
Z ⊕ Z ⊕
A= Ax dm and U AU = ∗
Atˆ dtˆ.
X b
G
The fact that these families are only defined up to a set of measure zero will be no
inconvenience.
Remark 1. It is always possible to choose the Ax such that
Tt∗ Aαt (x) Tt = Ax
holds for all x ∈ X and all t ∈ G. This can be seen in the following way: Theo-
rem 2.1.7 implies G ×α L∞ (X) ⊂ {Tt ⊗ S−t | t ∈ G}0 . In particular, we have for
fixed t ∈ G
Tt Aαt (x) Tt∗ = Ax , a.e. x ∈ X.
Therefore we get, using the Fubini theorem, that the family of operators defined by
fx ξ | ηi := M(t 7→ hTt Aαt x Tt∗ ξ | ηi),
hA
where M is the mean on the Abelian group G, coincides almost everywhere with
fx has the required
the family Ax . Moreover, it is easy to see that the family A
invariance property.
206 DANIEL H. LENZ
2.2. THE TRACE ON G ×α L∞ (X)
In the last section we proved that G ×α L∞ (X) is generated by a Hilbert algebra.

This allows to introduce a canonical trace on G ×α L∞ (X).
We start with a simple lemma that will allow us to prove the equality of certain
weights by proving the equality of the restrictions of these weights to suitable sets.
LEMMA 2.2.1. Let J be a strongly dense ideal in a von Neumann algebra N ⊂

B(H ) containing the identity I of B(H ).
(a) There is an increasing net Iλ in J converging strongly towards I . If H is
separable, (Iλ ) can be chosen as a sequence.
(b) If τ1 and τ2 are normal weights on N with τ1 (AA∗ ) = τ2 (AA∗ ) < ∞ for
A ∈ J, then τ1 = τ2 on (JJ)+ .
(c) If τ1 and τ2 are normal weights on N , whose restrictions to (JJ)+ coincide,
then τ1 = τ2 .
Proof. (a) The Ideal JJ is norm dense in the C ∗ algebra C := JJ. By 1.7.2
in [17], there exists therefore an approximate unit Iλ in JJ for C. As the net Iλ is
bounded and increasing, it converges strongly to some E ∈ B(H ) with
EC = CE = C, C ∈ C.
As J is strongly dense in N , the algebra C is weakly dense in N and EC = CE =

C, C ∈ N , follows. This implies E = I . If H is separable it is possible to choose
an increasing subsequence of (Iλ ) converging to E.
(b) This follows using polarisation.
(c) Let (Iλ ) be as in (a). Fix A = CC ∗ in N + . As τ1 and τ2 are normal, it is
enough to show
τ1 (CIλ C ∗ ) = τ2 (CIλ C ∗ ).
But this is clear, as JJ is an ideal and as (Iλ ) belongs to (JJ)+ . 2
THEOREM 2.2.2. There exists a unique normal trace τ on G ×α L∞ (X) with
τ (AA∗ ) = ha | ai
for A with kernel a. The trace τ is semifinite and τ = τc ◦ AdW ∗ , where τc is the
canonical trace on L(A) (cf. 2, Ch. 6, I of [18]). Moreover
(G ×α L∞ (X))2τ := {A ∈ G ×α L∞ (X) : τ (AA∗ ) < ∞} = K.
Proof. Clearly the identity of B(L2 (G × X)) is contained in the von Neumann
algebra G ×α L∞ (X) = π(G ×α C0 (X))00 and we can apply the foregoing lemma
with J = K to get the uniqueness.
As AdW is an isomorphism by Theorem 2.1.4, the remaining statements follow
easily from the corresponding statements in 2, Ch. 6, I of [18]. 2
DEFINITION 2.2.3. For a selfadjoint operator A affiliated to the von Neumann

algebra G ×α L∞ (X), i.e. whose resolution of the identity, EA [35], is contained
in G ×α L∞ (X), define
µA (B) := τ (EA (B))
for Borel measurable B ⊂ R. The map µ is called the integrated density of states
(IDS) for A (cf. [6]).
We mention that there is a different approach to the IDS for one-dimensional

Schrödinger operators due to Johnson and Moser (cf. [16, 24]).
From Theorem 2.2.2 we get the following corollaries.
COROLLARY 2.2.4. Let A and µA be as in the preceding definition. Then µA is

a spectral measure for A.
Proof. This is clear, as τ is faithful and normal. 2
COROLLARY 2.2.5. Let A and µ be as in Corollary 2.2.4. If there exists a set

σ ⊂ R with σ (Ax ) = σ a.e. x ∈ X, then σ = supp µ.
Proof. As µ is a spectral measure of A, we have σ (A) = supp µ. By σ (Ax ) = σ
a.e. x ∈ X, the equality σ = σ (A) holds. 2
There is another way to calculate the trace that can be seen as an application of
[13] (cf. Remark 1 below).
LEMMA 2.2.6. Let A be in (G ×α L∞ (X))+ . Then there exists a unique 3(A) ∈

[0, ∞] with
Z Z
3(A) g 2 (t) dt = tr(Mg Ax Mg ) dm
G X
for positive g ∈ L∞ (G), where tr denotes the usual trace on B(L2 (G)).
Proof. Uniqueness is obvious. Existence will follow from the uniqueness of the
Haar measure on G, once we have shown that the RHS of the equation induces an
invariant measure.
As A is positive there exists C ∈ G ×α L∞ (X) with A = C ∗ C. We calculate
Z Z
µ(B) := tr(MχB Ax MχB ) dm = tr(MχB Cx∗ Cx MχB ) dm.
X X
R
Using that tr is a trace we conclude µ(B) = X tr(Cx MχB Cx∗ ) dm.
This formula and some simple monotone convergence arguments show that µ
is a measure with
Z Z Z
g 2 dµ = tr(Cx Mg 2 Cx∗ ) dm = tr(Mg Ax Mg ) dm.
G X X
208 DANIEL H. LENZ
It remains to show that µ is translation invariant. As G ×α L∞ (X) is contained in

{Tt ⊗ S−t : t ∈ G}0 for each t ∈ G, the equation
Cαt (x) = Tt∗ Cx Tt
holds for a.e. x ∈ X. This allows to calculate
Z
µ(B − t) = tr(Cx MχB−t Cx∗ ) dm
ZX
= tr(Cx Tt MχB Tt∗ Cx∗ ) dm
X
Z
(tr is trace ) = tr(Tt∗ Cx Tt MχB Tt∗ Cx∗ Tt ) dm
ZX
= tr(Cαt (x) MχB Cα∗t (x)) dm
X
= µ(B).
Here we used in the last equation that m is translation invariant. The calculation
shows that µ is translation invariant. This finishes the proof. 2
THEOREM 2.2.7. 3 = τ .
Proof. As K is a strongly dense ideal in G ×α L∞ (X) by Proposition 2.1.6, it
is by Lemma 2.2.1 enough toR show 3(A∗ A) = τ (A∗ A) for A ∈ K. Choosing
a positive g ∈ L∞ (G) with G g 2 dt = 1, we calculate for A ∈ K with kernel
a ∈ L2 (G × X)
Z
∗
3(A A) = tr(Mg A∗x Ax Mg ) dm
Z Z
X
= |g(t)a(t − s, αt (x))|2 dt ds dm
Z
X G×G
Z Z !
(Fubini) = |g(t)|2 |a(t − s, αt (x))|2 dm ds dt
G G X
Z Z Z !
(m, ds transl. inv. ) = |g(t)|2
|a(s, x)| dm ds dt,
2
G G X
(kgkL2 (G) = 1) = ha | ai,
R an operator2K ∈ B(L (G)) with kernel k ∈ L (G × G) the

2 2
where we used that for
∗
equality tr(KK ) = G×G |k(t, s)| dt ds holds. 2
In some cases (e.g. in the almost periodic case or if G = Rn , Zn ) it is known

that there exists a sequence Hn ⊂ G with
Z Z
1
lim χHn (s)f (αs (x)) dmG (s) = f (z) dm(z)
n→∞ mG (Hn ) G X
for f ∈ L1 (X, m) and x ∈ Xf , where Xf is a suitable subset of XR with µ(X −

Xf ) = 0. If this is valid, and if A ∈ K has a kernel a s.t. x 7→ G |a(t, x)|2 dt
belongs to L1 (X, m), then
1
lim tr(χHn Ax A∗x χHn ) = τ (AA∗ ), (∗)
n→∞ mG (Hn )
R R
by tr(χHn Ax A∗x χHn ) = G G χHn (t)|a(s, α−t (x))|2 ds dt. Here, one can interpret a
term of the form χH Bx χH as the restriction Bx |H of Bx to L2 (H, mG |H ). One is
in particular interested in the case, where Ax = χI (Bx ) belongs to the resolution
of the identity of Bx . As restrictions of operators are comparatively accessible, the
question arises whether
1
lim tr(χHn χI (Bx )χHn ) − tr(χI (Bx |Hn )) = 0
n→∞ mG (Hn )
for x ∈ X and I ⊂ R. If this can be established, the equation

1
µA (I ) = lim tr(χI (Bx |Hn ))
n→∞ mG (Hn )
follows from (∗) immediately, as the equation χI (C) = χI (C)χI (C)∗ holds for
arbitrary operators C. Note that tr(χI (Bx |Hn )) is essentially the number of eigen-
values of Bx |Hn in I . For I = (−∞, E], E ∈ R, this type of equation has been
established for pseudodifferential operators with almost periodic coefficients in
[36], for Schrödinger operators in [6] using heat equation methods and for discrete
G in [3]. It is called Shubin’s trace formula.
Remarks. (1) In [13] transverse functions and transverse measures on groupoids
are introduced and studied (cf. I, 5 of [14] as well). It is possible to give G × X the
structure of a groupoid. The measure m then induces a unique transverse measure
3 with certain properties. It is possible to show that 3 satisfies the equation
Z Z
3(ξ ) f dt = ξx (f ) dm
G X
for transverse functions ξ and f ∈ L∞ (G). A direct calculation shows that for
A ∈ (G ×α L∞ (X))+ the mapping ξ x (B) := tr(χB Ax χB ) is a transverse function
(if the components Ax are chosen according to Remark 1 in Section 2.1). In this
context Theorem 2.1.7 says essentially 3(ξA ) = τ (A).
(2) In [1] it is shown that for a family Aω , ω ∈ , of almost periodic Schrödinger
operators and F ∈ C0 (R) the equation
Z Z
tr(Mg F (Aω )Mg ) dm = F dk
R
holds, where the measure dk is given by a certain limit procedure. Using Definition
2.2.3 and Theorem 2.2.7 we see dk = dµA .
210 DANIEL H. LENZ
2.3. SOME SPECIAL CASES
If m(X) < ∞ (e.g. if X is compact) there exist two alternative formulas for the
trace on G ×α L∞ (X). They will now be discussed.
Define for A ∈ G ×α L∞ (X) the operator Am : L2 (G) → L2 (G) by
Z
hAm ξ | ηi := hAx ξ | ηi dm, ξ, η ∈ L2 (G).
X
Since Aαt x = Tt Ax Tt∗ a.e. x ∈ X for fixed t ∈ G and m is invariant under α,

b) with
the operator Am is translation invariant. Therefore there exists ϕ ∈ L∞ (G
−1
Am = F Mϕ F , where Mϕ denotes the operator of multiplication by ϕ. Now it is
easy to see that the mapping
b A 7→ FAm F −1
J : G ×α L∞ (X) → M(L∞ (G)),
is linear, positive and faithfulR on (G ×α L∞ (X))+ . Let τ∞ be the usual trace on
b i.e. τ∞ (Mϕ ) := b ϕ dtˆ.
M(L∞ (G)), G
Moreover, define µ: (G ×α L∞ (X))+ → [0, ∞] by
Z
µ(A) := hAtˆ1 | 1i dtˆ,
b
G
where 1 denotes the function of L2 (X) with constant value 1. Then the following
holds.
THEOREM 2.3.1. τ = µ = τ∞ ◦ J .
Proof. We will show (1) τ = τ∞ ◦ J and (2) τ = µ.
(1) τ = τ∞ ◦ J : By Theorem 2.2.7, it is enough to show 3(A) = τ∞ ◦ J (A)
for A ∈ (G ×α L∞ (X))+ . For such an A let Am and the function φ be defined as
above, i.e. Am = F −1 Mφ F and Mφ = J (A). Choosing a positive g ∈ L∞ (G) with
kgkL2 (G) = 1 we calculate
Z
3(A) = tr(Mg Ax Mg ) dm = tr(Mg Am Mg )
X
= tr(Mg F −1 Mφ1/2 F F −1 Mφ1/2 F Mg ).
For φ ∈ L1 (G) b the operator K = Mg F −1 Mφ1/2 F is a Hilbert–Schmidt opera-
Rtor with kernel k(t, x) = g(t)F −1 (φ 1/2 )(t − s). Thus the formula tr(K ∗ K) =
G×G |k(t, s)| dt ds holds and we get
2
Z Z Z
−1
3(A) = |g(t)F (φ )(t − s)| dt ds = kφ k 2 b =
1/2 2 1/2
φ(tˆ) dtˆ,
G G
L (G)
Gb
where we used the translation invariance of dt, kgkL2 (G) = 1, and the fact that the
Fourier transform is an isometry. R
For arbitrary φ the equality 3(A) = G b φ(tˆ) dtˆ now follows by a simple
monotone limit procedure.
(2) τ = µ: For A ∈ (G ×α L∞ (X))+ let Am and φ be as above. By (1), it is

b But this follows from the
enough to show φ(tˆ) = hAtˆ1 | 1i =: ψ(tˆ) for a.e. tˆ ∈ G.
b):
following calculation valid for all η, ξ ∈ L2 (G
hMφ ξ | ηi = hAm F −1 ξ | F −1 ηi
Z
= hAx F −1 ξ | F −1 ηi dm
X
= hA(I ⊗ F −1 )(1 ⊗ ξ ) | (I ⊗ F −1 )(1 ⊗ η)i
= hAU ∗ (1 ⊗ ξ ) | U ∗ (1 ⊗ η)i
= hÂ(1 ⊗ ξ ) | (1 ⊗ η)i
= hMψ ξ | ηi.
The theorem is proven. 2
Remarks. (1) The expression µ was used in [25] (cf. also [9]).
(2) The mapping J was first introduced by Coburn, Moyer and Singer [10] (cf.
also [36]) in their paper on almost periodic operators.
(3) If X consists of only one point it is a forteriori compact. The positive oper-
ators in G ×α C0 (X) are just the operators A = FG−1 Mϕ FG where ϕ ∈ C0 (G) b is
positive (cf. Remark 1 in Section 1). The trace of such an A is then given by
Z
τ (A) = ϕ(tˆ) dtˆ,
b
G
b As the ideal of trace class opera-

in particular τ (A) is finite iff ϕ belongs to L1 (G).
tors consists of the finite linear combinations of positive operators with finite trace,
we conclude that A = FG−1 Mϕ FG is trace class iff ϕ belongs to L1 (G) b ∩ C0 (G).
b
In particular, it is not true in general that an operator of the form A = π(ϕ) =
F −1 MF (ϕ) F with ϕ ∈ Cc (G × X) = Cc (G) is trace class. This shows that it is not
possible (as is sometimes done) to define a trace on G ×α L∞ (X) by setting
Z
τ (π(ψ)) := ψ(0, x) dm.
X
We close this section with some remarks on the special case that G is discrete. As
in this case the function δe : G → C defined by δe (t) = 1 for t = e and δe (t) = 0
for t 6= e is positive, bounded with kδe kL2 (G) = 1, we get easily
Z
τ (A) = hAx δe | δe i dm.
X
Moreover, it is possible to show that there exists a conditional expectation
J : G ×α L∞ (X) → L∞ (X)
212 DANIEL H. LENZ
with τ = τ∞ ◦ J , where τ∞ is the usual trace on L∞ (X) (cf. [31] for a thorough
discussion of this case).
3. Groups Acting on Groups

This section is devoted to the study of a group acting on another group. This sit-
uation arises in particular in the context of almost periodic functions (cf. [8, 30])
and this is indeed the motivating example. The irrational rotation algebras, which
have recieved a lot of attention (cf. [15, 34] and references therein), arise in such a
situation. They are used in the treatment of the so called Harper’s model [4]. The
corresponding operator is just the almost Mathieu operator for λ = 2 (cf. [22] and
references therein for details about the almost Mathieu operator).
3.1. THE GENERAL CASE
We will look at the following situation: Let G and X be locally compact Abelian
groups, and let j : G → X be a group homomorphism. This induces a homomor-
b → G,
phism j ∗ : X b where Xb (resp. G)
b denotes the dual group of X (resp. G). Then
there is an action α of G on X given by
αt : X → X, αt (x) := x + j (t),
b on G
and an action of X b given by
b → G,
α̂ x̂ : G b α̂ x̂ (tˆ) := tˆ + j ∗ (x̂).
Similarly to the unitaries Tt and St resp., acting on L2 (G) and L2 (X) resp. for
t ∈ G, there are unitaries Tx̂ and Sx̂ defined by
b → L2 (G),
Sx̂ : L2 (G) b Sx̂ ξ(tˆ) := ξ(α̂ (−x̂) (tˆ)),
b → L2 (X),
Tx̂ : L2 (X) b Tx̂ ξ(ŷ) := ξ(ŷ − x̂).
Moreover, we define
W b × G)
b : L2 (X b → L2 (X
b × G)
b by W
b ξ(x̂, tˆ) := ξ(x̂, α̂ x̂ (tˆ))
e by
and U
e := (FX ⊗ FG )W ∗ = (FX ⊗ I )U,
U
where FX and FG are the Fourier transfom on L2 (X) and L2 (G) resp. and U and
W are as defined in the first section.
We will first establish a spatial isomorphism between the von Neumann algebras
G ×α L∞ (X) and X b ×α̂ L∞ (Gb). This can be done quite easily, as the generators
of these von Neumann algebras are known explicitly.
We will then provide proofs for some statements first appearing in [6] and [7],
that yield much more, namely an isomorphism between G ×α C0 (X) and X b ×α̂
b).
C0 (G
We will need the following propositions.
PROPOSITION 3.1.1. (a) W ∗ (Tt ⊗I ) = (Tt ⊗St )W ∗ , W b ∗ (Tx̂ ⊗I ) = (Tx̂ ⊗Sx̂ )W ∗ .
(b) W (Tt ⊗ I ) = (Tt ⊗ S−t )W, Wb (Tx̂ ⊗ I ) = (Tx̂ ⊗ S−x̂ )W .
∗
e b
(c) U W = W U . e
Proof. This can be seen by direct calculation. 2
PROPOSITION 3.1.2. (a) The von Neumann algebra G ×α L∞ (X) is generated

by operators of the form Tt ⊗ I , t ∈ G, and the operators of multiplication with the
functions
(t, x) 7→ (x̂ | αt (x)) = (x̂ | x)(x̂ | j (t)), x̂ ∈ X̂.
(b) The von Neumann algebra X b is generated by operators of the
b ×α̂ L∞ (G)
form Tx̂ ⊗ I, t ∈ G, and the operators of multiplication with the functions
(x̂, tˆ) 7→ (t | α̂ x̂ (tˆ)) = (t | j ∗ (x))(t | tˆ).
Proof. This follows from Theorem 2.1.7 and the well known fact that the char-
acters generate the von Neumann algebra L∞ . 2
Now we can prove the spatial isomorphism, mentioned above.

THEOREM 3.1.3. The unitary U e: L2 (G×X) → L2 (X b × G)
b establishes a spatial
∞
isomorphism between G ×α L (X) and X b ×α̂ L (G).
∞ b
Proof. By the foregoing proposition, it is enough to show
e(Tt ⊗ I )U
(1) U e∗ = M((t | j ∗ (x))(t | tˆ)),
e e∗ = Tx̂ ⊗ I ,
(2) U M((x̂ | x)(x̂ | j (t)))U
where M(ϕ) denotes as usual the operator of multiplication by ϕ.
(1) Using Proposition 3.1, we can calculate
e(Tt ⊗ I ) = (FX ⊗ FG )W ∗ (Tt ⊗ I )
U
= (FX ⊗ FG )(Tt ⊗ St )W ∗
= (t | ·)(t | j ∗ ·)(FX ⊗ FG )W ∗ .
(2) This can be seen by similar arguments. 2
As already stated above there is even an isomorphism between the C ∗ -algebras

G ×α C0 (X) and X b ×α̂ C0 (G).
b We will establish this isomorphism in two steps.
In the first step we will show that the image of an operator with L2 -kernel under
conjugation by U e is also an operator with an L2 -kernel. In the second step we will
prove that conjugation by U e is actually an isomorphism between G ×α C0 (X) and
214 DANIEL H. LENZ
b ×α̂ C0 (G).
X b We remark that both of these facts have already been stated in [6]
and [7], where, however, no proof was given.
LEMMA 3.1.4. Let a ∈ L2 (G × X) be the kernel of a bounded operator A on
L2 (G × X), i.e.
Z
Aξ(t, x) = a(t − s, αs (x))ξ(s, x) ds.
G
e := U
Then A eAU is a bounded operator on L2 (X
e∗ b × G)
b with kernel
b × G),
â := (FX ⊗ FG )W a ∈ L2 (X b
i.e.
Z
e x̂, tˆ) =
Aξ( â(x̂ − ŷ, α̂ x̂ (tˆ))ξ(ŷ, tˆ) dŷ.
X̂
b × G).
Proof. Let ξ̂ := (FX ⊗ FG )(ξ ) be an arbitrary function in L2 (X b We
calculate
eAU
U e∗ ξ̂(x̂, tˆ)
= (FX ⊗ FG )W ∗ AW ξ(x̂, tˆ)
Z !
= (FX ⊗ FG ) (t, x) 7→ a(t − s, x)W ξ(s, α−t (x)) ds (x̂, tˆ)
G
Z !
= (FX ⊗ FG ) (t, x) 7→ (W a)(t − s, αt −s (x))ξ(s, αt −s (x)) ds (x̂, tˆ).
G
R
As for fixed t ∈ G the mapping x 7→ G |(W a)(t − s, αt −s (x))ξ(s, αt −s (x))| ds
belongs to L1 (X), this expression equals
Z Z ! !
(I ⊗ FG ) (x̂ | −x) (W a)(t − s, αt −s (x))ξ(s, αt −s (x)) ds dx (x̂, tˆ),
X G
which yields after the substitution (x 7→ x − j (t − s))

· · · = (FX ⊗ FG )((t, x) 7→ ((x̂ | −j (·))W a)(·, x) ∗ ξ(·, x)(t))(x̂, tˆ)
= FX (x 7→ (I ⊗ FG )(W a)(tˆ + j ∗ (x̂), x)(I ⊗ FG )(ξ )(tˆ, x))(x̂)
= (FX ⊗ FG )(W a)(tˆ + j ∗ (x̂), ·) ∗ (FX ⊗ FG )(ξ )(tˆ, ·)(x̂)
Z
= â(x̂ − ŷ, α̂ x̂ (tˆ))ξ̂(ŷ, tˆ) dŷ.
X
This proves the lemma. 2
b ×α̂ C0 (G
THEOREM 3.1.5. The mapping AdUe : G ×α C0 (X) → X b), A 7→ U
eAU
e∗
∗
is an isomorphism of C -algebras.
Proof. We have to show that AdUe (G ×α C0 (X)) is contained in X b ×α̂ C0 (G)

b
b ×α̂ C0 (G))
and that AdUe∗ (X b is a subset of G ×α C0 (X). We show
(1) There is a dense set D ⊂ G ×α C0 (X) with AdUe (D) ⊂ X b ×α̂ C0 (G).
b
b b
(2) There is a dense subset F of X ×α̂ C0 (G) with AdUe∗ (F ) ⊂ G ×α C0 (X).
(1) Let
D := {π(W ∗ (g ⊗ (h1 ∗ h2 ))) | g ∈ Cc (G), h1 , h2 ∈ Cc (X)}.
It is easy to see that D is in fact dense in G ×α C0 (X). The lemma yields that for
A = π(W ∗ (g ⊗ (h1 ∗ h2 ))) ∈ D the operator AdUe (A) has the kernel
â := FG (g) ⊗ (FX (h1 )FX (h2 )).

But, as FG (g) belongs to C0 (G) b and (FX (h1 )FX (h2 )) belongs to L1 (X),
b the
function â is indeed the kernel of an operator in G ×α C0 (X).
(2) This can be seen by similar arguments. 2
We now provide a proof for another theorem which was already stated (without
proof) in [6] and [7].
THEOREM 3.1.6. Let τ (resp. τ̂ ) be the trace on G ×α L∞ (X) (resp. X b ×α̂
∞ b e e ∗
L (G)) defined in the last section. Then the equation τ (A) = τ̂ (U AU ) holds
for all A ∈ (G ×α L∞ (X))+ .
Proof. It is enough to consider the case A = BB ∗ with B ∈ K with kernel b
(cf. Lemma 2.2.1). Then the kernel of U eB Ue∗ is given by (FG ⊗ FX )(W b) and we
have
Z Z
eAU
τ̂ (U e∗ ) = |(FG ⊗ FX )(W b)(x̂, tˆ)|2 dx̂ dtˆ
b
G b
X
Z Z
= |W b(t, x)|2 dt dx
G X
Z Z
= |b(t, x)|2 dt dx
X G
= τ (BB ∗ ).
As A = BB ∗ the theorem is proven. 2
There is an analogue of the classical Plancherel Theorem.
COROLLARY 3.1.7. The mapping AdUe establishes an isomorphism between the

b ×α̂ L∞ (G))
ideals (G ×α L∞ (X))2τ and (X b b2τ with
τ (AA∗ ) = b
τ (AdŨ (A)AdŨ (A)∗ ).
216 DANIEL H. LENZ
Proof. This follows directly from Theorem 3.1.6. 2
We will now give a short application of the above theory.
3.2. PERIODIC OPERATORS
Let H be a closed subgroup of a locally compact Abelian group G such that X :=

G/H is compact. Let p: G → X be the canonical projection. Then the machinery
developed in the last section can be applied with j = p. For A ∈ G ×α L∞ (X) we
b 3 tˆ 7→ Atˆ the family of operators with
denote in this section by G
Z ⊕
eAU
Atˆ dtˆ = U e∗ ,
b
G
e was defined in the last section.

where U
The operators in G ×α L∞ (X) have a very strong invariance property.
∞
PROPOSITION R ⊕ 3.2.1. For every A ∈ G ×∗
α L (X) there exist unique Ax , x ∈ X,
with (i) A = X Ax dm and (ii) Tt Aαt (x)Tt = Ax , x ∈ X, t ∈ G.
The same holds for selfadjoint A that are affiliated to G ×α L∞ (X).
Proof. For A ∈ G ×α L∞ (X) the existence of such Ax has already been shown
(cf. Remark 1 in Section 2). The uniqueness follows as p is surjective. For self-
adjoint A affiliated to G ×α L∞ (X) the uniqueness proof is unchanged. Existence
follows by looking at (A + i)−1 . 2
We have the following theorem.

THEOREM 3.2.2. Let A be selfadjoint and affiliated to G ×α L∞ (X) with res-
olution of identity EA and fibres Ax , x ∈ X, chosen according to the preceding
proposition. Then the measure µ defined in Definition 2.2.3 is a spectral measure
for all x ∈ X.
Proof. By Corollary 2.2.4, the measure µ is a spectral measure for A. As all Ax ,
x ∈ X, are unitarily equivalent by Proposition 3.2.1, the statement follows. 2
Remarks. (1) Kaminker and Xia show in [25] by the use of a spectral duality
principle that certain elliptic periodic operators have purely continuous spectra on
the complement of the set of discontinuities of λ 7→ τ (EA ((−∞, λ])). Theorem
3.2.2 shows in particular that this holds for arbitrary periodic operators for purely
algebraic reasons.
(2) For periodic Schrödinger operators it is possible to show that the spectrum
is purely absolutely continuous using some analyticity arguments (cf. [33] and
references therein).
We finish this section with a short discussion of another formula for τ . Let
H⊥ ⊂ Gb be the annihilator of H and let q: Gb → G/Hb ⊥ denote the canonical
projection. For g = f ◦ q define {g} := f . Let for ρ = q(tˆ) the functional

b → R be defined by
Iρ : L∞ (G)
Z X
Iρ (f ) := f (tˆ + h⊥ ) dmH ⊥ = f (tˆ + h⊥ ),
H⊥
h∈H ⊥
then the desintegration formula

Z Z
ˆ ˆ
f (t ) dt = Iρ (f ) dmG/H
b ⊥ (ρ)
b
G b ⊥
G/H
holds (cf. [20]).

Let A ∈ (G ×α L∞ (X))+ be given. Identifying (G/H \) with H ⊥ and using
X
tr Atˆ = hAtˆδh | δh i = Iq(tˆ) (ŝ 7→ hAŝ δe | δe i),
h∈H ⊥
we calculate
Z
τ (A) = hAtˆδe | δe i dtˆ
b
ZG
= Iρ (ŝ 7→ hAŝ δe | δe i) dmG/H

b ⊥ (ρ)
b ⊥
G/H
Z
= {ŝ 7→ tr Aŝ }(ρ) dmG/H
b ⊥ (ρ).
b ⊥
G/H
The RHS of this equation is essentially the integrated density of states defined
in Ch. XIII of [33] for periodic operators.
4. Spectral Duality
By spectral duality we mean a relation between the spectral types of Ax , x ∈ X,
b of the form “If Atˆ has pure point spectrum a.e.
and the spectral types of Atˆ, tˆ ∈ G,
b, then Ax has purely (absolutely) continuous spectrum a.e. x ∈ X”.
tˆ ∈ G
Theorems of this form have been stated in [7, 9, 25]. We cite the theorem of [25].
THEOREM 4.1. Let A ∈ G ×α L∞ (X) be selfadjoint with purely continuous
spectrum on a Borel set E s.t. Atˆ has pure point spectrum on E for almost all
b then Ax has purely continuous spectrum on E for almost all x ∈ X.
tˆ ∈ G,
In [22] another form of duality is proven for the Almost Mathieu Equation. The
method developed there can be carried over with only small changes to give
THEOREM 4.2. Let (Z, α, X, m) be a dynamical system and assume that X is a
complete metric space. Let A ∈ Z×α L∞ (X) be selfadjoint with spectral family EA
s.t. Ax has only pure point spectrum with simple eigenvalues for almost all x ∈ X.
Then µ(B) := τ (EA (B)) is a spectral measure for Aη for almost all η ∈ bZ =: S 1 .
218 DANIEL H. LENZ
Proof. As µ is a spectral measure for A by Corollary 2.2.4, it is enough to

show that there are spectral measures ν η for Aη not depending on η. This is shown
following [22].
By the same method as in [22], it can be shown that there exist measurable
functions
Nj : X → N ∪ {0} ∪ {∞}, and ϕjl : X → l 2 (Z), j ∈ Z, l ∈ N,
s.t. ϕjl (x) = 0 for all j ∈ Z, x ∈ X, l ∈ N with Nj (x) < l, the set
{ϕjl (x) | j ∈ Z, l = 1, . . . , Nj (x)}
is an orthonormal basis of l 2 (Z) consisting of eigenvectors of Ax for almost all

x ∈ X and that the ϕjl satisfy
ϕjl (x) = Tk ϕjl −k (αk (x)), k ∈ Z.

In particular, we have
hϕjl (x) | Tk ϕjl (αk (x))i`2 (Z) = 0, k 6= 0,
for all x ∈ X as either ϕjl (x) and Tk ϕjl (αk (x)) = ϕjl +k (x) are different members of
an ONB or at least one of them is zero.
The simplicity of the eigenvalues is crucial to get this measurable section of
eigenfunctions. We will now show (cf. [22])
(1) Fix ψ ∈ L2 (X), F ∈ C0 (R), j ∈ Z, l ∈ N. Let ξ(z, x) := ψ(x)ϕjl (x)(z),
ξ̂η (x) := U ξ(η, x) and
µη (F ) := hξ̂η | F (Aη )ξ̂η iL2 (X) .
Then µη (F ) is independent of η a.e. η.
(2) µη does not depend on η a.e. η.
(3) There exist νη not depending on η s.t. νη is a spectral measure for Aη for a.e. η.
By the remarks at the beginning
R of the proof, the theorem follows from (3).
(1) It is enough to show 0 = S 1 (η | z)µη (F ) dη for all z ∈ Z with z 6= 0. We
calculate
Z
(η | z)µη (F ) dη = h(I ⊗ Mz )U ξ | U F (A)ξ i
S1
= hU (Tz ⊗ Sz )ξ | U F (A)ξ i
= h(Tz ⊗ Sz )ξ | F (A)ξ i.
Using that ϕjl (x) is an eigenvector corresponding to the eigenvalue ejl (x) say, we
get
Z
··· = F (ejl (x))ψ(αz (x))ψ(x)hTz ϕjl (αz (x)) | ϕjl (x)i dm
X
= 0,
where we used the relation hTz ϕjl (αz (x)) | ϕjl (x)i = 0.
(2) As C0 (R) is separable, this follows from (1).
(3) Let {ψm } be an ONB of L2 (X). Then {ψm ⊗ ϕjl } is an ONB in L2 (Z × X)
and, as U is unitary, it follows that the ξl,j,m := U (ψm ⊗ ϕjl ) form an ONB in
L2 (Ẑ × X). Thus the set T := {ξl,j,m (η, ·) | l, j, m} is total in L2 (X) for almost all
η ∈ S 1 . (Notice that the set Mϕ := {η ∈ S 1 | ϕ ⊥ ξl,j,m (η, ·)∀l, j, m} has measure
zero for each ϕ ∈ L2 (G).) Therefore the measures ν η defined by
X
ν η (B) := hχB (Aη )ξl,j,m (η, ·) | ξl,j,m (η, ·)iL2 (X)
l,j,n
are spectral measures for almost all η ∈ S 1 , which do not depend on η by (2). The
theorem follows. 2
Acknowledgements
The author would like to thank P. Stollmann for many useful discussions. Financial
support from Studienstiftung des deutschen Volkes (Doktorandenstipendium) is
gratefully acknowledged.
References
1. Avron, J. and Simon, B.: Almost periodic Schrödinger Operators, II. The integrated density of
states, Duke Math. J. 50 (1982), 369–391.
2. Bellissard, J.: Schrödinger operators with an almost periodic potential, in R. Schrader and T.
Seiler (eds), Mathematical Problems in Theoretical Physics, Springer, Berlin, 1982, pp. 356–
363.
3. Bellissard, J.: K-Theory of C ∗ -algebras in solid state physics, in T. C. Dorlas, N. M. Hugenholz
and M. Winnik (eds), Statistical Mechanics and Field Theory: Mathematical Aspects, Lecture
Notes in Phys. 257, Springer, Berlin, 1986, pp. 99–156.
4. Bellissard, J.: Gap labelling theorems for Schrödinger operators, in M. Waldschmitt, P. Moussa,
J. M. Luck and C. Itzykson (eds), From Number Theory to Physics, Springer, New York, 1992,
pp. 538–630.
5. Bellissard, J., Bovier, A. and Ghez, J.-M.: Gap labelling theorems for Schödinger operators,
Rev. Math. Phys. 4 (1992), 1–37.
6. Bellissard, J., Lima, R. and Testard, D.: Almost periodic Schrödinger operators, in L. Streit
(ed.), Mathematics + Physics, World Scientific, Singapore, 1985, pp. 1–64.
7. Bellissard, J. and Testard, D.: Quasi-periodic Hamiltonians. A mathematical approach, Proc.
Sympos. Pure Math. 38 (1982), 579–582.
8. Bohr, H.: Almost Periodic Functions, Chelsea Publishing Company, 1947.
9. Chojnacki, W.: A generalized spectral duality theorem, Comm. Math. Phys. 143 (1992), 527–
544.
10. Coburn, L. A., Moyer, R. D. and Singer, I. M.: C ∗ -algebras of almost periodic pseudo-
differential operators, Acta Math. 139 (1973), 279–307.
11. Combes, J. M. and Hislop, P. D.: Localization for some continuous, random Hamiltonians in
d-dimensions, J. Funct. Anal. 124 (1994), 149–180.
12. Combes, J. M., Hislop, P. D. and Mourre, E.: Spectral averaging, perturbation of singular
spectra and localization, Trans. Amer. Math. Soc. 348 (1996), 4883–4894.
13. Connes, A.: Sur la théorie non commutative de l’intégration, Lecture Notes in Math. 725,
Springer, Berlin, 1979.
220 DANIEL H. LENZ
14. Connes, A.: Noncommutative Geometry, Academic Press, San Diego, 1994.
15. Davidson, K. R.: C ∗ -Algebras by Example, Fields Institute Monographs, 1996.
16. Delyon, F. and Souillard, B.: The rotation number for finite difference operators and its
properties, Comm. Math. Phys. 89 (1983), 415–426.
17. Dixmier, J.: C ∗ -Algebras, North-Holland, Amsterdam, 1975.
18. Dixmier, J.: Von Neumann Algebras, North-Holland, Amsterdam, 1981.
19. van Elst, A.: Gap-labelling theorems for Schrödinger operators on the square and cubic lattice,
Rev. Math. Phys. 6 (1994), 319–342.
20. Folland, G. B.: A Course in Abstract Harmonic Analysis, Stud. Adv. Math., CRC Press, Bocca
Raton, 1995.
21. Forrest, A., Hunton, J. and Kellendonk, J.: Projection quasicrystals I: Toral rotations, Preprint.
22. Gordon, A., Jitomirskaya, S., Last, Y. and Simon, B.: Duality and singular continuous spectrum
in the almost Mathieu equation, Acta Math. 178 (1997), 169–183.
23. Haagerup, U.: On the dual weights for crossed products of von Neumann algebras. I. Removing
separability conditions; II Application of operator valued weights, Math. Scand. 43 (1978),
99–118, 119–140.
24. Johnson, R. and Moser, J.: The rotation number for almost periodic potentials, Comm. Math.
Phys. 84 (1982), 403–438.
25. Kaminker, J. and Xia, J.: The spectrum of operators elliptic along the orbits of R n actions,
Comm. Math. Phys. 110 (1987), 427–438.
26. Kellendonk, J.: Nonncommutative geometry of tilings and gap labelling, Rev. Math. Phys. 7
(1995), 1133–1180.
27. Kellendonk, J.: The local structure of tilings and their integer group of coinvariants, Comm.
Math. Phys. 187 (1997), 115–157.
28. Kirsch, W., Stollmann, P. and Stolz, G.: Localization for random perturbations of periodic
Schrödinger operators, Random Operators Stochastic Equations 6 (1998), 241–268.
29. Klopp, F.: Localization for some continuous random Schrödinger operator, Comm. Math. Phys.
167 (1995), 553–569.
30. Maak, W.: Fastperiodische Funktionen, Springer, Berlin, 1950.
31. Pedersen, G. K.: C ∗ -Algebras and Their Automorphism Groups, Academic Press, London,
1987.
32. Pimsner, M. and Voiculescu, D.: Exact sequences for K-groups and Ext groups of certain cross
product C ∗ -algebras, J. Operator Theory 4 (1980), 93–118.
33. Reed, M. and Simon, B.: Analysis of Operators IV, Academic Press, San Diego, 1980.
34. Rieffel, M.: C ∗ -algebras associated with irrational rotations, Pacific J. Math. 93 (1981), 415–
429.
35. Rudin, W.: Functional Analysis, Tata McGraw-Hill, New Delhi, 1974.
36. Shubin, M. A.: The spectral theory and the index of elliptic operators with almost periodic
coefficients, Russian Math. Surveys 34 (1979), 109–157.
37. Takesaki, M.: Tomita’s Theory of Modular Hilbert Algebras and Its Applications, Lecture Notes
in Math. 128, Springer, New York, 1970.
38. Takesaki, M.: Theory of Operator Algebras, I, Springer, New York, 1979.
39. Tomiyama, J.: Invitation to C ∗ -Algebras and Topological Dynamics, World Scientific, 1987.
40. Tomiyama, J.: The Interplay between Topological Dynamics and Theory of C ∗ -Algebras, Lec-
ture Note Series, Vol. 2, Research Institute of Mathematics, Global Analysis Research Center,
Seoul National University, 1992.
41. Tomiyama, J.: C ∗ -algebras and topological dynamical systems, Rev. Math. Phys. 8 (1996),
741–760.
42. Weidmann, J.: Linear Operators in Hilbert Spaces, Grad. Texts in Math. 68, Springer, New
York, 1980.
223
Schrödinger Operators with Empty Singularly

Continuous Spectra
MICHAEL DEMUTH
Institute of Mathematics, Technical University of Clausthal, 38678, Clausthal-Zellerfeld, Germany
KALYAN B. SINHA?
Indian Statistical Institute, New Delhi, India
(Received: 4 May 1999)
Abstract. Let H be a semibounded perturbation of the Laplacian H0 in L2 (Rd ). For an admissible

function ϕ sufficient conditions are given for the completeness of the scattering system {ϕ(H ),
ϕ(H0 )}. If ϕ is the exponential function and if e−λH is an integral operator we denote the kernel
of the difference Dλ = e−λH − e−λH0 by Dλ (x, y), λ > 0. The singularly continuous spectrum of
H is empty if
Z Z
dx dy|Dλ (x, y)|(1 + |y|2 )α < ∞
Rd Rd
for some α > 1. This result is applied to potential perturbations and to perturbations by imposing
Dirichlet boundary conditions.
Mathematics Subject Classifications (1991): 47A10, 47A40, 81Q10, 81U99.
Key words: singularly continuous spectrum, Schrödinger operators, obstacle scattering.
1. Introduction
The main objective of this article is to prove the absence of the singularly continu-
ous spectrum for self-adjoint operators H in L2 (Rd ) which are perturbations of the
Laplacian H0 = −1. The proofs are based on the theory of complete scattering
systems. Instead of studying H0 and H directly we investigate scattering systems
for functions of H0 and H , respectively. To this aim, the abstract time-dependent
completeness theory is generalized to such functions. In particular, the function
t → e−λt is admissible. In many cases, the corresponding semigroups are integral
operators, the kernels of which can be represented by the Feynman–Kac formula
in terms of the conditional Wiener measure. Using this, the abstract operator theo-
retical conditions can be simplified. For potential perturbations and for the case of
obstacle scattering, one gets explicit conditions for the potential and the obstacle
? Research partly supported by Jawaharlal Nehru Centre for Advanced Scientific Research,
Bangalore.
224 MICHAEL DEMUTH AND KALYAN B. SINHA
region. The potential should decrease like (1 + |x|2 )− 2 −ε , ε > 0, as usual, or

1
satisfy an integral condition. For singular perturbations unbounded obstacles are

included.
2. Results
Some of the following ideas go back to Enss [5 – 7]. His proof of completeness was
oriented towards studying potential perturbations of the Laplacian. Here we have
mainly the obstacle scattering in mind where the potentials are arbitrarily high over
certain regions in Rd . For that we transform the problem of the completeness of the
scattering system {H, H0 } to the question of whether the system {ϕ(H ), ϕ(H0)} is
complete, where ϕ is a real-valued function with positive or negative first deriv-
ative. The abstract operator theoretical conditions in Theorem 1 ensure that the
singularly continuous subspace of H vanishes, i.e. the singularly continuous spec-
trum is empty. In Corollary 4, these general conditions are translated to semigroups,
i.e. to ϕ(H ) = e−λH . If the semigroups are ultracontractive one gets integral
conditions for their kernels which guarantee the absence of singularly continuous
spectrum (see Theorem 5).
The present section gives the results which will be proved in Section 3. In Sec-
tion 4, we apply the theory to potential perturbations and to singular perturbations
obtained by imposing Dirichlet conditions on the obstacle region. It turns out that
the following conditions are sufficient: For potential scattering (see Theorem 13,
Remark 14)
|V (x)| 6 c(1 + |x|2 )− 2 −ε ,

1
for all |x| > 1, and some ε > 0,
or
Z
dx(1 + |x|2 )α |V (x)| < ∞, with some α > 1.
Rd
For scattering by an obstacle, the conditions are given in terms of the equilibrium
potential v0 (·) (cf. (79)) of the singularity region 0 or in terms of its capacity (see
(81) and
R Theorem 15, Corollary 16). The theory can be applied if cap(0) is finite
and if dxv0 (x)(1 + |x|2 )α < ∞ for some α > 1. This allows unbounded 0 (see
Example 17).
We introduce the following notation. Let H0 be the self-adjoint realization of
the Laplacian in L2 (Rd ). Let H be an arbitrary self-adjoint operator in L2 (Rd ) and
assume H > −1. For the sake of abbreviation, we sometimes use H := L2 (Rd ).
Let ϕ be a real-valued function in C ∞ (R). We assume either ϕ 0 (λ) > 0 or ϕ 0 (λ) <
0 for all λ ∈ σc (H0 ) ∪ σc (H ), where σc (·) denotes the continuous spectrum of the
corresponding operator. Functions ϕ with these properties are called admissible.
SCHRÖDINGER OPERATORS WITH EMPTY SINGULARLY CONTINUOUS SPECTRA 225
Denote by ζ1 (·), ζ2 (·) continuous functions from σ (H ) ∪ σ (H0) → R \ {0} such

that 1/ζ1 , 1/ζ2 exist and are bounded. Also we shall use the letter c as a generic
constant in all estimates.
THEOREM 1. Let H0 , H, ϕ, ζi be given as explained above. Additionally, we

assume that
(i)
ζ2−1 (H )[ϕ(H ) − ϕ(H0 )]ζ1−1 (H0 ) ∈ B∞ (H),
i.e., the sandwiched difference is a compact operator on H = L2 (Rd ).
(ii) Assume that
e−λH − e−λH0 ∈ B∞ (H) (1)
for all λ ∈ R+ , or assume that
(H + 1)−1 − (H0 + 1)−1 ∈ B∞ (H). (2)
(iii) Define the multiplication operator by
(M|x|2 f )(x) := |x|2 · f (x) (3)
for x ∈ Rd and f in dom(M|x|2 ), given by
dom(M|x|2 ) := {f ∈ H : | · |2 f (·) ∈ H}.
Assume that for some ε > 0
ζ2−1 (H )[ϕ(H ) − ϕ(H0 )]ζ1−1 (H0 )(1I + M|x|2 ) 2 +ε ∈ B(H),
1
(4)
i.e., it is a bounded operator on H = L2 (Rd ).
Then the singularly continuous subspace of H vanishes:
Hsc (H ) = {θ}, (5)
where θ is the zero vector, i.e., the singularly continuous part of the spectrum is
empty
σsc (H ) = ∅. (6)
In fact, we prove somewhat more. Assumptions (i)–(iii) imply the existence and
completeness of the wave operators
± (ϕ(H ), ϕ(H0)) = s − lim eit ϕ(H ) e−it ϕ(H0 ) Pac (ϕ(H0 )), (7)
t →±∞
where Pac (ϕ(H0 )) is the projection operator onto the absolutely continuous sub-
space of the self-adjoint operator ϕ(H0 ); here Pac (ϕ(H0))H = H. We prove
ran[± (ϕ(H ), ϕ(H0))] = Hc (ϕ(H )), (8)
i.e., the range of + and the range of − are equal to the continuous subspace of
ϕ(H ).
Remark 2. H and ϕ(H ) have the same spectral measure. Hence, Hc (ϕ(H )) =
Hc (H ). On the other hand, Hac (ϕ(H )) = Hac (H ) because ϕ 0 6= 0 almost every-
where (see [10], p. 518, Example 1.9). On the other hand,
ran[+ (ϕ(H ), ϕ(H0))] = ran[− (ϕ(H ), ϕ(H0))] ⊆ Hac (ϕ(H ))
⊆ Hc (ϕ(H )).
This implies
Hac (H ) = Hac (ϕ(H )) = Hc (ϕ(H )) = Hc (H )
or
Hsc (H ) = {θ}.
Thus, the existence of the wave operators in (7) and the identity (8) imply the
assertions in (5) and (6).
Remark 3. The admissible functions in the present paper form a subclass of

functions for which the invariance principle can be formulated (see, e.g., [2], p. 158).
If ϕ is admissible, ϕ −1 is also admissible. The composition of two admissible
functions is again admissible (cf. [8]), The invariance principle in its strong sense
implies the existence of ± (H, H0 ) if ± (ϕ(H ), ϕ(H0)) exist and their equality.
Sufficient conditions on ϕ such that the invariance principle holds in its strong
sense can be derived from [13], p. 49, Theorem XI. 23. For such functions ϕ,
Theorem 1 also gives the existence and completeness of ± (H, H0 ).
COROLLARY 4. Take as an admissible function ϕ(σ ) = e−σ , and take

1 1
= = e−σ .
ζ1 (σ ) ζ2 (σ )
Assume
e−λH − e−λH0 ∈ B∞ (H), ∀λ ∈ R+ , (9)
and
e−H (e−H − e−H0 )e−H0 (1I + M|x|2 ) 2 +ε ∈ B(H),
1
ε > 0. (10)
Then ± (e−H , e−H0 ) exist and are complete.
Moreover, we have
+ (e−H , e−H0 ) = − (H, H0 ),
− (e−H , e−H0 ) = + (H, H0 ),
and
Hsc (H ) = {2},
i.e., σsc (H ) = ∅.
THEOREM 5. Let λ > 0 be fixed but arbitrary and let e−λH be an integral
operator in L2 (Rd ) with the kernel (e−λH )(x, y) admitting an estimate of the form
|x−y|2
|e−λH (x, y)| 6 cλ− 2 e−γ
d
λ ecλ (11)
with 0 < γ < 14 .
Assume, furthermore, that the kernel Dλ (x, y) of the integral operator
Dλ := e−λH − e−λH0 (12)
satisfies one of the following conditions: Either
Z
sup |x|α |Dλ (x, y)| dy < ∞ (13)
|x|>R >1 2|y|>R
with some α > 1 , or

Z Z
dx dy|Dλ (x, y)|(1 + |y|2 )α < ∞ (14)
Rd Rd
for some α > 1.
Then Theorem 1 and Corollary 4 are applicable, i.e., σsc (H ) = ∅.
3. Proofs
3.1. PROOF OF THEOREM 1
The proof of Theorem 1 is split into several lemmata.

We will prove Theorem 1 only for ϕ 0 < 0. The case ϕ 0 > 0 is similar and is
omitted.
For the proof we need a special spectral decomposition of the Hilbert space.
For that, see also [12], p. 74 ff, and [15]. One has to find two associated projection
operators P+ and P− such that
P+ + P− = 1I, (15)
where P+ and P− have some further properties to be specified later. First we define
P+ , P− :
DEFINITION 6. Let F be the Fourier transform from L2 (Rd , dx) to L2 (Rd , dk).
Let U be a transformation from L2 (Rd , dk) to L2 (R, db; L2 (S d−1 )) given by
1 d b
(U Ff )b (ω) = √ e 4 b (Ff )(e 2 ω), (16)
2
ω ∈ S d−1 , where S d−1 is the unit sphere of Rd and L2 (R, b; L2 (S d−1 )) denotes the
space of square-integrable L2 (S d−1 )-valued functions on R.
(Note that U and F are unitary. For a smooth function ψ(·): R → R we obtain
(U F ψ(H0 )f )b (ω) = ψ(eb )(U Ff )b (ω), f ∈ dom(ψ(H0 )),
i.e., U F maps into the spectral representation in which ln(H0 ) is diagonal.)
Let G be the one-dimensional Fourier transform with respect to b, i.e.,
Z
1
(GU Ff )a (ω) := √ e−iab (U Ff )b (ω) db. (17)
2π R
G maps L2 (R, db; L2 (S d−1 )) to L2 (R, da; L2 (S d−1 )).
Now we define the operators P+ , P− by
P+ f := F ∗ U ∗ G∗ Mχ(a>0) GU Ff, (18)
P− f := F ∗ U ∗ G∗ Mχ(a<0) GU Ff. (19)
Here (Mg f )(x) := g(x)f (x) is the multiplication operator by the function g (see
(3)) and χ(a>0) denotes the characteristic function of the set {a : a > 0}.
Clearly, P+ , P− are orthogonal projection operators satisfying (15).
Similarly, we define an operator Pαt on L2 (Rd ) for α > 0 by
Pαt f := F ∗ U ∗ G∗ Mχ(|a|6αt )GU Ff. (20)
LEMMA 7. Assume H0 , H, ϕ as in Section 2 with ϕ 0 < 0. Assume (i), (ii), (iii) of

Theorem 1. Assume the existence of ± (ϕ(H ), ϕ(H0)) (see (7)).
Let P+ , P− , Pαt be as in Definiton 6. Furthermore, let ψ ∈ Cc∞ (R) such that
0∈/ supp(ψ), and that for an appropriate α > 0
kPαt exp{i|t|ϕ(H0 )}ψ(H0 )P+ k 6 c(1 + |t|)−l (21)
and
kPαt exp{−i|t|ϕ(H0 )}ψ(H0 )P− k 6 c(1 + |t|)−l , (22)
where l is a positive integer that can be arbitrarily large and where the constants
c depend only on ϕ and l.
Then
ζ2−1 (H )[+(ϕ(H ), ϕ(H0)) − 1I]ψ(H0 )P− ∈ B∞ (H), (23)
ζ2−1 (H )[−(ϕ(H ), ϕ(H0)) − 1I]ψ(H0 )P+ ∈ B∞ (H), (24)
s − lim P+ e−it ϕ(H0 ) = 0, (25)

t →∞
s − lim P− eit ϕ(H0 ) = 0. (26)

t →∞
COROLLARY 8. Suppose the assumptions and the conclusion of Lemma 7 are

valid, then
ran(± ) = Pc (ϕ(H ))H = Pc (H )H. (27)
Hence Hsc (H ) = {θ} and σsc (H ) = ∅, where we have used the abbreviation
± = ± (ϕ(H ), ϕ(H0)) for short.
First we prove the last corollary, then Lemma 7. Finally, we prove that (21),
(22) are valid.
Proof of Corollary 8.
Preremark. Assumption (ii) in Theorem 1 is only used to show that the dif-
ference ψ(H ) − ψ(H0 ) is a compact operator in H for any function ψ(·) ∈
Cc∞ (R).
Proof. Set
Z T
1
ε+ (g(·)) := lim g(t) dt (28)
T →∞ T 0
if the right-hand side exists. Set

D := {ψ̃(H )g, g ∈ Hc (H ), ψ̃ ∈ Cc∞ (R \ {0})}. (29)
D is dense in Hc (H ) (see, e.g., [2], p. 331). By this definition, one can find for any
vector f ∈ D a real-valued function ψ ∈ Cc∞ (R \ {0}) such that f = ψ(H )f .
Of course, the choice of this ψ depends on f . Note also that since ± exist by
hypothesis, ± intertwines H and H0 , i.e. ψ(H )± = ± ψ(H0 ).
For any f ∈ D we have the following identity
kf k2 = ε+ (e−iϕ(H )· f, + ψ(H0 )P− e−iϕ(H )· f ) − (30)
− ε+ (e−iϕ(H )· f, ψ(H )(+ − 1I)ψ(H0 )P− e−iϕ(H )· f ) + (31)
+ ε+ (e−iϕ(H )· f, − ψ(H0 )P+ e−iϕ(H )· f ) − (32)
− ε+ (e−iϕ(H )· f, ψ(H )(− − 1I)ψ(H0 )P+ e−iϕ(H )· f ) + (33)
+ ε+ (e−iϕ(H )· f, [ψ(H ) − ψ(H0 )]e−iϕ(H )· f ). (34)
Now
ψ(H )ζ2 (H )ζ2−1 (H )(+ − 1I)ψ(H0 )P− and
ψ(H )ζ2 (H )ζ2−1 (H )(− − 1I)ψ(H0 )P+
are compact by (23) and (24), while ψ(H ) − ψ(H0 ) is compact for any ψ ∈
Cc∞ (R \ {0}) as per the Preremark. Therefore, the expressions in (31), (33), and
(34) are zero using Wiener’s Theorem or the derived RAGE Theorem (see, e.g.,
[12], p. 23 ff). The term in (32) vanishes because of the intertwining relation and
(25). Finally, one has
kf k2 = ε+ (e−iϕ(H )· f, + ψ(H0 )P− e−iϕ(H )· f ). (35)
Since D is dense, the identity in (35) is true for any f ∈ Hc (ϕ(H )). Now, if
ran(+ ) ⊂ Hc (ϕ(H )) there would be an f0 , f0 6= θ, with f0 ⊥ ran(+ ), i.e.
∗+ f0 = 0. Then the intertwining property of the wave operator + together with
(35) would imply f0 = θ. This is a contradiction. Hence, ran(+ ) = Hc (ϕ(H )). 2
For proving ran(− ) = Hc (H ), one has to choose an analogous decomposition

of kf k2 .
This proves Corollary 8. 2
Proof of Lemma 7. First we prove (25) and (26). For that we define the positive
self-adjoint operator A2 in L2 (Rd ) by
A2 f := F ∗ U ∗ G∗ Ma 2 GU Ff, (36)
which is densely defined with
dom(A2 ) := {f ∈ L2 (Rd ), A2 f ∈ L2 (Rd )}.
Then, by (21),
k(1 + A2 )−1 ei|t |ϕ(H0 ) ψ(H0 )P+ k (37)
6 k(1 + A2 )−1 (1 − Pαt )k + k(1 + A2 )−1 Pαt ei|t |ϕ(H0 ) ψ(H0 )P+ k
1
6c + ckPαt ei|t |ϕ(H0 ) ψ(H0 )P+ k
1 + α2t 2
1
6c .
1 + α2t 2
Hence,
lim kP+ ψ(H0 )e−i|t |ϕ(H0 ) (1 + A2 )−1 k = 0. (38)
t →∞
The range of (1 + A2 )−1 is dense in H by the definition of A2 and (38) is true for
any ψ ∈ Cc∞ (R). The set {ran[ψ(H0 )(1 + A2 )−1 ], ψ ∈ Cc∞ } is dense in H. Thus,
(21) and (38) imply
s − lim P+ e−it ϕ(H0 ) = 0.
t →∞
Similarly, (26) will follow from (22).
Now we show the compactness of ζ2−1 (H )(− − 1I)ψ(H0 )P+ . For the moment,
the wave operator is assumed to exist (the proof is given in Lemma 10). Thus, we
have
ζ2−1 (H )(− (ϕ(H ), ϕ(H0)) − 1I)ψ(H0 )P+
Z 0
= −i ds eisϕ(H ) ζ2−1 (H )[ϕ(H ) − ϕ(H0 )]ζ1−1 (H0 )e−isϕ(H0 ) ζ1 (H0 )ψ(H0 )P+ .
−∞
From Assumption (i) we know that ζ2−1 (H )[ϕ(H )−ϕ(H0)]ζ1−1 (H0 ) is compact.
Therefore the assertion is true if
Z ∞
kζ2−1 (H )(ϕ(H ) − ϕ(H0 ))ζ1−1 (H0 )e−isϕ(H0 ) ζ1 (H0 )ψ(H0 )P+ k ds
0
is finite.
For that, we use the decomposition
kζ2−1 (H )(ϕ(H ) − ϕ(H0 ))ζ1−1 (H0 )e−isϕ(H0 ) ζ1 (H0 )ψ(H0 )P+ k

6 kζ2−1 (H )(ϕ(H ) − ϕ(H0 ))ζ1−1 (H0 )(1 + M|x|2 ) 2 +ε k ×
1
(39)
− 12 −ε −1 1 +ε
× k(1 + M|x|2 ) (H0 + 1) (1 + A ) k × 2 2

× k(1 + A2 )− 2 −ε (1I − Pαs )e−isϕ(H0 ) ζ1 (H0 )(H0 + 1)ψ(H0 )P+ k +
1

+ k(1 + A2 )− 2 −ε Pαs e−isϕ(H0 ) ζ1 (H0 )(H0 + 1)ψ(H0 )P+ k .
1
The first factor is bounded by Assumption (iii). From Definition 6, it easily

follows that the operator A given by Ma in the spectral representation GU F H
is precisely the self-adjoint extension of the generator of the group of dilations
in L2 (Rd ) whose restriction to S(Rd ) is −(i/4)(x · 5 + 5 · x). Therefore, the
boundedness of the second factor in (39) follows from [11], Lemma 6.1, Asser-
tion (xx). The first term in the curly brackets decreases like (1 + s 2 )− 2 −ε , since
1
ζ1 (H0 )(H0 + 1)ψ(H0 ) is bounded, while in the second term in the curly brackets
ζ1 (·)(· + 1)ψ(·) is again in Cc∞ (R \ {0}) such that the estimate in (21) is applicable.
Finally, (23) follows from (22) in the same way. 2
Now we will show that the inequality in (21) holds for an appropriate α given
the assumptions in Theorem 1.
LEMMA 9. Let P+ , P− , Pαt be given as in Definition 6. Fix a function ψ(·) ∈

Cc∞ (R) with 0 ∈
/ supp ψ. Then there is an α > 0 such that
kPαt eit ϕ(H0 ) ψ(H0 )P+ k 6 c(1 + |t|)−l , for all l ∈ N.
Proof. Define
ρ(H0 ) = eit ϕ(H0 ) ψ(H0 ).
Then
hPαt ρ(H0 )P+ f, f 0 iL2 (Rd ) (40)

= hF ∗ U ∗ G∗ Mχ(|a|6αt )G Mρ(eb ) G∗ Mχ(a 0 >0) GU Ff, f 0 i
= hχ(| · | 6 αt)G Mρ(eb ) G∗ Mχ(a 0 >0) g, g 0 iL2 (R,da;L2(S d−1 )) ,
where we have set g = GU Ff , and g 0 = GU Ff 0 . Thus, the above is equal to

Z Z
dω daχ(|a| 6 αt)[G Mρ(eb ) G∗ χ+ g](a, ω)g 0 (a, ω)
S d−1
Z R Z Z
= dω da da 0 χ(|a| 6 αt)χ(a 0 > 0)K(a, a 0 )g(a 0 , ω)g 0 (a, ω)
S d−1 R R
with
Z
1 0
K(a, a 0 ) = eib(a−a ) ρ(eb ) db. (41)
2π R
Hence,
|hPαt ρ(H0 )P+ f, f 0 i|

Z Z 1/2
0 0 2
6 da da |K(a, a )| χ(|a|6αt )χ(a 0>0) kf kkf 0 k.
R R
Therefore,
kPαt eit ϕ(H0 ) ψ(H0 )P+ k2 (42)

Z Z Z 2
1
da da χ(|a|6αt )χ(a 0 >0) ρ(e ) db .
0 0)
6 eib(a−a b
R R 2π R
We now study
Z
0 b
eib(a−a ) eit ϕ(e ) ψ(eb ) db (43)
R
for |a| 6 αt and a 0 > 0. Using integration by parts, the last integral is equal to
∞
ψ(eb )
−ie i[b(a−a 0 )+t ϕ(eb )] , (44)
(a − a 0 ) + teb ϕ 0 (eb ) −∞
Z ∞
1 0 b )] d ψ(eb )
− db ei[b(a−a )+t ϕ(e . (45)
i −∞ db a − a 0 + teb ϕ 0 (eb )
The term in (44) is zero because ψ(e−∞ ) = ψ(0) = 0 and ψ(∞) = 0, and because
the denominator in (44) is bounded, which will be shown next.
We are in the case ϕ 0 < 0. Thus, there are two positive constants ã1 , ã2 with
−ã1 < ϕ 0 (eb ) < −ã2
as long as eb ∈ supp ψ. Hence, for t > 0 there are two positive constants a1 , a2
such that
−ta1 < tϕ 0 (eb )eb < −ta2 .

For a 0 > 0, it follows that
−a 0 + tϕ 0 (eb )eb < −a 0 − ta2 .
Choosing α < a2 , we obtain
|a − a 0 + teb ϕ 0 (eb )| > t (a2 − α). (46)
Thus, the term in (44) vanishes. For estimating the term in (45), we repeat this
procedure l-times. Then
c 1
|K(a, a 0 )| 6 . (47)
t l−2 |a − a 0 + teb ϕ 0 (eb )|2
This implies
Z Z ∞
c
da da 0 |K(a, a 0 )|2 6 . (48)
|a|6αt 0 t 2l−2
Here l is arbitrary because ϕ is assumed to be in C ∞ (R). The proof of (22) is
similar and this completes the proof of Lemma 9. 2
In order to complete the proof of Theorem 1, it remains to show the existence

of ± (ϕ(H ), ϕ(H0)).
LEMMA 10. Under the assumptions (i), (ii), (iii) of Theorem 1, the wave opera-
tors ± (ϕ(H ), ϕ(H0)) exist.
Proof. Again we consider only the case ϕ 0 < 0 and the existence of + . It
suffices to prove the existence on a dense set. For that we choose
{ψ(H0 )f, f ∈ dom(A2 ), ψ(·) ∈ Cc∞ (R \ {0})}.
For any ψ1 (·) ∈ Cc∞ (R \ {0}), [ψ1 (H ) − ψ1 (H0 )] is compact and therefore we have
the identity
s − lim eit ϕ(H ) e−it ϕ(H0 ) ψ1 (H0 ) = s − lim ψ1 (H )eit ϕ(H ) e−it ϕ(H0 ) (49)
t →∞ t →∞
if one of these strong limits exist. We show the existence of the second limit. By
restricting to a dense set again, it is sufficient to establish that
s − lim ψ1 (H )eit ϕ(H ) e−it ϕ(H0 ) ψ2 (H0 )

t →∞
exists on dom(A2 ), where ψ2 (·) is another arbitrary function in Cc∞ (R \ {0}).

Hence, it suffices to show that
Z ∞
kψ1 (H )[ϕ(H ) − ϕ(H0 )]e−it ϕ(H0 ) ψ2 (H0 )f k dt < ∞ (50)
tf
for f ∈ dom(1I + A2 ). For this, we estimate as in the proof of Lemma 7:

kψ1 (H )ζ2 (H )ζ2 (H )−1 [ϕ(H ) − ϕ(H0 )]ζ1−1 (H0 )e−it ϕ(H0 ) ζ1 (H0 )ψ2 (H0 )f k
6 kψ1 (H )ζ2 (H )k kζ2 (H )−1 [ϕ(H ) − ϕ(H0 )]ζ1−1 (H0 )(1 + M|x|2 ) 2 +ε k ×
1
× k(1 + M|x|2 )− 2 −ε (H0 + 1)−1 (1 + A2 ) 2 +ε k ×

1 1
× k(1 + A2 )− 2 −ε e−it ϕ(H0 ) (H0 + 1)ζ1 (H0 )ψ2 (H0 )f k.

1
The first factor is bounded since ψ1 ∈ Cc∞ , the second is assumed to be bounded
in Assumption (iii), the third is bounded according to [11] (see (39)). It remains to
estimate the t-dependence of the last factor.
Because
c
k(1 + A2 )− 2 −ε (1I − Pβ|t | )k 6 1+ε for each β > 0,
1
t
it is enough to estimate
kPβ|t | e−it ϕ(H0 ) (H0 + 1)ζ1 (H0 )ψ2 (H0 )f k.
Because f = (1 + A2 )−1 g, we consider
kPβ|t | e−it ϕ(H0 ) (H0 + 1)ζ1 (H0 )ψ2 (H0 )(1 + A2 )−1 k.
For α > 0
c
k(1I − Pα|t | )(1 + A2 )−1 k 6 ,
t2
thus the proof will be complete if we can get an integrable estimate for
kPβ|t | e−it ϕ(H0 ) (H0 + 1)ζ1 (H0 )ψ2 (H0 )Pα|t | k

with a suitable choice of α and β.
As in Lemma 7 (see (42)), the integral kernel here is
Z
0
db eib(a−a ) χ(|a| 6 βt)χ(|a 0 | 6 αt)e−it ϕ(e ) (eb + 1)ζ1 (eb )ψ2 (eb ).
b
Integration by parts produces here the denominator a − a 0 − teb ϕ 0 (eb ). For suitable
α, β, this is bounded from below by
|a − a 0 − teb ϕ 0 (eb )| > c|t|,
which implies
kPβ|t | e−it ϕ(H0 ) (H0 + 1)ζ1 (H0 )ψ2 (H0 )Pα|t | k 6 cl · |t|−l , l ∈ N.
This proves Lemma 10 and the proof of Theorem 1 is complete. 2

3.2. PROOF OF THEOREM 5
According to Corollary 4, one has to show that for λ ∈ (0, ∞)
e−λH − e−λH0 ∈ B∞ (H) (51)
and that
K := e−H (e−H − e−H0 )e−H0 (1I + M|x|2 ) 2 +ε ∈ B(H).

1
(52)
(K is defined at first on dom (1I + M|x|2 ) 2 +ε and then extended.)

1
LEMMA 11. The operator K,
K = e−H (e−H − e−H0 )e−H0 (1I + M|x|2 ) 2 +ε

1
(53)
= e−H D1 e−H0 (1I + M|x|2 )
1 +ε
2 ,
is an integral operator with the kernel

Z Z
dv(e−H )(x, u)D1 (u, v)(e−H0 )(v, y)(1 + |y|2 ) 2 +ε .(54)
1
K(x, y) = du
Rd Rd
K is bounded if for some R > 1 and ε > 0

Z
sup |x| 1+ε
|D1 (x, y)| dy < ∞ (55)
|x|>R >1 2|y|>R
or if
Z Z
dx dy |D1 (x, y)|(1 + |y|2 )1+ε < ∞. (56)
Rd Rd
Proof. The operator norm of K can be estimated by

Z Z
kKk 6 sup
2
|K(x, y)| dy · sup |K(x, y)| dx. (57)
x Rd y Rd
For the first factor, we obtain

Z Z Z
dv |e−H (x, u)||D1 (u, v)|e−H0 (v, y)(1 + |y|2 ) 2 +ε (58)
1
sup dy du
R
ZR ZR
x d d d
dv |e−H (x, u)||D1 (u, v)|(1 + |v|2 ) 2 +ε .

1
6 c sup du (59)
x Rd Rd
The last estimate follows from the standard result:

Z
|x−y|2 d
e− 2λ (1 + |y|2 )α dy 6 c (1 + |x|2 )α λ 2 (1 + λα ), λ > 0. (60)
Rd
The term in (59) is smaller than

Z Z
dv |e−H (x, u)||D1 (u, v)|(1 + |v|2 ) 2 +ε +
1
c sup du
x Rd |v|62|u|
Z Z
(|v|−|u|)2
dv |e−H (x, u)| e− 8 (1 + |v 2 |) 2 +ε
1
+ c sup du
x Rd |v|>2|u|
Z Z
2 12 +ε
6 c sup du (1 + |u| ) dv |D1 (u, v)||e−H (x, u)| +
x Rd |v|62|u|
Z Z
|v|2
dv |e−H (x, u)|e− 32 (1 + |v 2 |) 2 +ε
1
+ c sup du
x Rd |u|6 |v|
Z 2
6 c sup(1 + |u|2 ) 2 +ε
1
|D1 (u, v)| dv + c. (61)
u |v|62|u|
For |u| 6 R, the supremum in (61) is always finite. For |u| > R > 1
Z
2 12 +ε
sup (1 + |u| ) |D1 (u, v)| dv (62)
|u|>R >1 Rd
Z
(|u|−|v|2
6 c sup |u| 1+2ε
e− 8 dv +
|u|>R |v|6 R2
Z
+ c sup |u|1+2ε |D1 (u, v)| dv.
|u|>R 2|v|>R
The first term in (62) is always finite, the second is assumed to be finite in (55).
The second factor in (57) can be estimated in a similar way. In brief, we have
Z Z Z
dv |e−H (x, u)||D1 (u, v)|e−H0 (v, y)(1 + |y|2 ) 2 +ε
1
sup dx du
y Rd Rd Rd
Z Z
dv |D1 (u, v)|e−H0 (v, y)(1 + |y|2 ) 2 +ε
1
6 c sup du
y Rd Rd
Z
e−H0 (v, y)(1 + |y|2 ) 2 +ε dv +
1
6 c sup
y |v|6 |y|
Z 2
Z
du |D1 (u, v)|(1 + |v|2 ) 2 +ε |e−H0 (v, y)|
1
+ c sup dv
y 2|v|>|y| Rd
|y|2
6 c sup(1 + |y|2 ) 2 +ε |y|d e−
1
16
y
Z
1 +ε
+ c sup(1 + |v| ) 2 2 |D1 (u, v)| du. (63)
v Rd
Hence, the condition in (55) implies that both factors on the left-hand side of (57)
are finite.
Condition (56) implies that K is a Hilbert–Schmidt operator. We find

Z Z
dx dy |K(x, y)|2 (64)
Z Z Z Z
2
= dx dy du dv e−H (x, u)D1 (u, v)e−H0 (v, y) (1 + |y|2 )1+2ε
Z Z
6 dx dy (1 + |y|2 )1+2ε ×
Z Z Z Z
× du1 dv1 du2 dv2 e−H (x, u1 )|D1 (u1 , v1 )|e−H0 (v1 , y) ×
× e−H (x, u2 )|D1 (u2 , v2 )|e−H0 (v2 , y)

Z Z Z Z
6 c dv1 du1 |D1 (u1 , v1 )| dv2 du2 |D1 (u2 , v2 )| ×
Z
× dy (1 + |y|2 )1+2ε e−H0 (v1 , y)e−H0 (v2 , y)
Z Z Z Z
6 c dv1 du1 |D1 (u1 , v1 )| dv2 du2 |D1 (u2 , v2 )| ×
Z
× dy (1 + |y|2 )1+2ε e−H0 (v1 , y)
Z Z 2
6c dv du |D1 (u, v)|(1 + |v| ) 2 1+2ε
, (65)
using the estimate in (60) with respect to the y-integration. 2
It remains to show the compactness in (51) to complete the proof of Theorem 5.

For that we formulate the next lemma.
LEMMA 12. The difference e−λH − e−λH0 is compact if for each β > 0
Z
sup |x|β
|Dλ (x, y)| dy < ∞ (66)
|x|>R >1 Rd
or if
Z Z
dx dy |Dλ/2 (x, y)| < ∞. (67)
Rd Rd
Proof. The operator

Mχ(|x|6n) Dλ Mχ(|y|62n)
n ∈ R+ , is a Hilbert–Schmidt operator with the Hilbert–Schmidt norm satisfying
Z Z 1/2
6 cλ− 4 nd .
d
dx dy |Dλ (x, y)|2
|x|6n |y|62n
It suffices to prove
Z
lim sup dy |χ(|x| 6 n)Dλ (x, y)χ{|y| 6 2n} − Dλ (x, y)| = 0, (68)
n→∞ x Rd
where one should keep in mind that

Z
sup dx |χ(|x| 6 n)|Dλ (x, y)χ(|y| 6 2n)|
y Rd
is uniformly bounded in n. For |x| 6 n, we have in (68)

Z
sup |Dλ (x, y)χ(|y| 6 2n) − Dλ (x, y)| dy
|x|6n Rd
Z
(|y|−|x|)2
dy λ− 2 e−c 8λ
d
6 sup
|x|6n |y|>2n
n2
6 cλ− 2 e−c λ
d
which tends to zero as n → ∞.

For the rest in (68), that means for |x| > n we have
Z Z
|x|β
sup |Dλ (x, y)| dy 6 sup β dy |Dλ (x, y)|
|x|>n |x|>n n
Z
1
6 β sup |x| β
dy |Dλ (x, y)|
n |x|>R Rd
which tends also to zero as n → ∞. This proves e−λH − e−λH0 to be compact if the
condition in (66) is satisfied.
For (67), we use the standard decomposition
e−λH − e−λH0 = e− 2 H (e− 2 H − e− 2 H0 ) + (e− 2 H − e− 2 H0 )e− 2 H0 .

λ λ λ λ λ λ
Both terms are Hilbert–Schmidt operators because, e.g.,

Z Z Z 2

dx dy du (e − λ2 H
)(x, u)Dλ/2 (u, y)
Z Z 2
cλ − d2
6 ce λ dy du |Dλ/2 (u, y)| .
Rd Rd
R
Because of (14) Rd |Dλ/2 (u, y)| du is finite, and this means that the condition (67)
is sufficient for e−λH − e−λH0 to be Hilbert–Schmidt. Lemma 12 is proved. 2
Remark. The proof of Theorem 5 is complete now. The conditions in (55) and
(56) are stronger than the ones in (66) or (67), respectively. Therefore, we need
either the condition in (13) or that in (14). As we will see in the applications,
these two conditions have a common intersection but generally allow different
perturbations. The dependence on λ in (13) or (14) is not critical.
4. Applications
4.1. POTENTIAL PERTURBATIONS
THEOREM 13. Let H0 = −1 be the Laplacian in L2 (Rd ). Assume a potential

V : Rd → R which is a Kato-class potential (a definition can be found, e.g., in the
book of Cycon et al. [3]). Furthermore assume that
sup |x|α |V (x)| < ∞ (69)

|x|>R >1
for some α > 1.

Then
σsc (H0 + MV ) = ∅.
Proof. Because V is a Kato-class potential, we know from the Feynman–Kac

formula
|x−y|2
(e−λ(H0 +MV )(x, y) 6 cλ− 2 e−
d
8λ ecλ ,
i.e., (11) is satisfied (see [14], Proposition B.6.7).

Next we show the condition in (13). By Duhamel’s formula (see, e.g., [3])
Z
sup |x| α
dy |(e−λ(H0 +MV ) )(x, y) − (e−λH0 )(x, y)|
|x|>R >1 Rd
Z Z Z λ
6 sup |x|α dy du ds |e−sH0 (x, u)| ×
|x|>R >1 Rd Rd 0
−(λ−s)(H0 +MV )
× |V (u)|e (u, y)|
Z Z λ
6 cecλ sup |x|α du ds |e−sH0 (x, u)||V (u)|.
|x|>R >1 Rd 0
For |u| 6 R/2, |x| > R we have

Z Z λ
du ds |(e−sH0 )(x, u)||V (u)|
|u|6 R2 0
Z Z λ
−(|x|−|u|)2
ds s − 2 e
d
6c du 4s |V (u)|
|u|6 R2 0
Z Z λ
d |x|2 R2
6c du ds s − 2 e−c λ e−c s |V (u)|
|u|6 R2 0
2
Z Z λ
R2
−c |x|λ
ds s − 2 e−c
d
6 ce du |V (u)| s
|u|6 R2 0
2
Z
−c |x|λ
6 cλe |V (u)| du.
|u|6 R2
The last term is finite because each Kato-class potential is in L1loc (Rd ).
For |u| > R/2 we obtain
Z Z λ
sup |x|α du ds |(e−sH0 )(x, u)|V (u)|
|x|>R |u|> R2 0
Z Z λ
d |x−u|2 |u|α
6 c sup |x| α
du ds s − 2 e− 4s |V (u)| +
|x|>R |x|
2 >|u|> 2
R
0 Rα
Z Z λ
|u|α
+ c sup |x|α du ds |(e−sH0 )(x, u)|V (u)|
|x|>R |u|> |x| 0 |x|α
Z Z
2
λ
d (|x|−|u|)2
6 cR −α sup |V (u)||u|α sup |x|α du ds s − 2 e− 4s +
|u|> R2 |x|>R |u|6 |x|
2 0
Z Z λ
+ c sup sup |u|α |V (u)| dv ds |(e−sH0 )(x, v)|
|x|>R |u|> R2 Rd 0
Z λ
|x|2 R2
6 cR −α sup (|V (u)||u|α ) · sup |x|α e−c ds s − 2 e−c
d
λ s
|u|> R2 |x|>R 0
+ cλ sup (|u| |V (u)|). α
|u|> R2
Remark 14. The condition in (69) or (13) is the usual one for short-range po-
tentials
1
|V (x)| 6 c , for |x| > R. (70)
(1 + |x|)1+ε
The condition in (14) is satisfied if
Z
dx (1 + |x|2 )α |V (x)| < ∞, (71)
Rd
which also allows unbounded potentials.
4.2. OBSTACLE SCATTERING
The Laplace operator H0 in L2 (Rd ) is associated to the Wiener process. Its semi-
group e−λH0 can be represented by
(e−λH0 f )(x) = Ex {f (X(λ))} (72)
for all f ∈ L2 (Rd ). Here Ex {·} is the expectation with respect to the Wiener
measure.
The obstacle region is denoted by 0, where 0 is a closed set in Rd . We denote
its complement by 6 = Rd \0. Define S0 := inf{s, X(s) ∈ 0} the first hitting time
of 0 for the underlying Wiener process. We assume that the set of regular points of
0 coincide with the regular points of the interior of 0. Setting
(Tλ f )(x) = Ex {f (X(λ)), S0 > λ}, (73)
we obtain a strongly continuous semigroup {Tλ , λ > 0} on L2 (6). The generator
of this semigroup is the Friedrichs extension of H0 restricted to Cc∞ (6) (see, e.g.,
[1]). Denoting this generator by H̃6 , we have
(e−λH̃6 f )(x) = Ex {f (X(λ)), S0 > λ}. (74)

This is a semigroup in L2 (6), which means that the scattering of the pair {H0 , H̃6 }
is in fact a two-space scattering problem. However, this can be reformulated as a
problem in the single Hilbert space L2 (Rd ) by the extension
e−λH6 := e−λH̃6 ⊕ 0
on L2 (6) ⊕ L2 (0). The difference of the two semigroups in L2 (Rd ) is given by
(e−λH0 f − e−λH6 f )(x) = Ex {[1 − χ{S0 > λ}]f (X(λ))}
= Ex {f (X(λ)), S0 < λ}. (75)
y,λ
Using the conditional Wiener measure Ex {·}, this semigroup difference is an
integral operator with the kernel
Dλ (x, y) = (e−λH0 )(x, y) − (e−λH6 )(x, y)
= Exy,λ {S0 < λ}. (76)
Note that this kernel is symmetric in x and y, and recall the crucial role in the
conditions of Theorem 5 played by the integral of |Dλ (x, y)| with respect to one of
the variables. With regard to the condition (14), we note that
Z
|Dλ (x, y)| dx = Ey {S0 < λ}. (77)
Rd
Hence, we obtain the following result.
THEOREM 15. Let H0 and H6 be given as described above.

Let
Z
Ex {S0 < λ}(1 + |x|2 )α dx < ∞ (78)
for some α > 1, λ > 0.

Then the wave operators ± (H6 , H0 ) exist and are complete. The singularly
continuous spectrum of H6 is empty, i.e., σsc (H6 ) = ∅.
Remark. The expectation in (78) can be estimated by the one-equilibrium po-

tential v0 of 0, defined as
v0 (x) := Ex {e−S0 , S0 < ∞} (79)
which maps Rd to [0, 1]. Obviously
Ex {S0 < λ} 6 eλ v0 (x). (80)
Moreover, the one-equilibrium potential is related to the capacity of the obstacle
set 0. In the present context, the capacity of 0 is given by
Z
cap(0) = v0 (x) dx (81)
Rd
(see, e.g., [4, 9]).

Using the Hölder’s inequality in (78), one finds that the assertion of Theorem
15 is true if
Z 1/q
cap(0) 1/p
Ex {S0 < λ}(1 + |x| ) }
2 αq
<∞ (82)
with (1/p) + (1/q) = 1.

For bounded 0, we have
Ex {S0 < λ} = 1 if x ∈ 0
and
[dist(x,0)]2
Ex {S0 < λ} 6 ce−c λ
for x ∈/ 0. This means that Theorem 15 is true for all bounded 0. Hence, we get
the following result.
COROLLARY 16. Let cap(0) < ∞ and assume

Z
v0 (x)(1 + |x|2 )δ dx < ∞
Rd
for some δ > 1. Then σsc (H6 ) = ∅.

The condition in (78) also allows unbounded 0 given in the next example.
EXAMPLE 17. Let the dimension d satisfy d > 3, let 0 be a union of balls
B(an , rn ) = Bn centered in an = (n, 0, 0, . . . , 0), i.e., |an | = n, and with radii
rn < 1.
Then Theorem 15 is true for 0 satisfying

X d−2
rn p n2α < ∞
n
with p > 1, α > 1. This 0 is in general an unbounded set.
Proof of the example. Corresponding to the sufficient condition in (78), we have

to estimate
XZ
Ex {SBn < λ}(1 + |x| ) dx
2 α
n Rd
XZ Z
6 dx (1 + |x|2 )α + Ex {SBn < λ}(1 + |x|2 )α dx}
n Bn Rd \B n
X Z
(|x−an |−rn )2
1/q
6c rn |an |2α + cap(Bn )1/p dx e−c λ (1 + |x|2αq )
n Rd \Bn
X d−2
6c {rn |an |2α + rn p (1 + |an |2α )}
n
X d−2
6c rn p n2α . 2
n
P 1 d−2−2α
In particular, one can take rn = (1/n)p , then n n is finite for d > 6 and
1 6 α < 1, 5.
References
1. Baumgärtel, H. and Demuth, M.: Decoupling by a projection, Rep. Math. Phys. 15 (1979),
173–186.
2. Baumgärtel, H. and Wollenberg, M.: Mathematical Scattering Theory, Birkhäuser, Basel, 1983.
3. Cycon, H. L., Froese, R. G., Kirsch W., and Simon, B.: Schrödinger Operators with Applica-
tions to Quantum Mechanics, Springer-Verlag, Berlin, 1987.
4. Demuth, M., McGillivray, I., and Noll, A.: Capacity and spectral theory, In: M. Demuth, E.
Schrohe, B.-W. Schulze and J. Sjöstrand (eds), Spectral Theory, Microlocal Analysis, Singular
Manifolds, Adv. in Partial Differential Equations 14, Akademie-Verlag, Berlin, 1997, pp. 12–
77.
5. Enss, V.: Asymptotic completeness for quantum-mechanical potential scattering I. Short range
potentials, Comm. Math. Phys. 61 (1978) 285–291.
6. Enss, V.: Asymptotic completeness for quantum-mechanical potential scattering II. Singular
and long-range potentials, Ann. Phys. 119 (1979), 117–132.
7. Enss, V.: Geometrical methods in spectral and scattering theory of Schrödinger operators, In:
G. Velo and A. S. Wightman (eds), Rigorous Atomic and Molecular Physics, Plenum Press,
New York, 1981.
8. Eder, S.: Factorization theorems for compact operators and semigroup differences, PhD Thesis,
TU Clausthal, 1999.
9. Fukushima, M., Oshima, Y., and Takeda, M.: Dirichlet Forms and Symmetric Markov
Processes, de Gruyter, Berlin, 1994.
10. Kato, T.: Perturbation Theory for Linear Operators, Springer-Verlag, Berlin, 1976.
11. Muthuramalingam, Pl. and Sinha, K. B.: Asymptotic completeness in long-range scattering II,
Ann. Sci. Ecole Norm. Sup. 4 18 (1985), 57–87.
12. Perry, P. A.: Scattering Theory by the Enss Method, Math. Rep. 1, Harwood Acad. Publ., Chur,
1983.
13. Reed, M. and Simon, B.: Methods of Modern Mathematical Physics, Vol. III: Scattering Theory,
Academic Press, New York, 1979.
14. Simon, B.: Schrödinger semigroups, Bull. Amer. Math. Soc. (N.S.) 7 (1982), 447–526.
15. Sinha, K. B.: Mathematical theory of scattering in quantum mechanics – a review, Current Sci.
56 (1987), 293–297.
245
An Asymptotic Expansion for Bloch Functions on

Riemann Surfaces of Infinite Genus and Almost
Periodicity of the Kadomcev–Petviashvilli Flow
FRANZ MERKL
Courant Institute of Mathematical Sciences, 251 Mercer Street, New York, NY 10012, U.S.A.
e-mail: merkl@cims.nyu.edu?
(Received: 28 June 1999)
Abstract. This article describes the solution of the Kadomcev–Petviashvilli equation with C 10 real
periodic initial data in terms of an asymptotic expansion of Bloch functions. The Bloch functions
are parametrized by the spectral variety of a heat equation (heat curves) with an external potential.
The mentioned spectral variety is a Riemann surface of in general infinite genus; the Kadomcev–
Petviashvilli flow is represented by a one-parameter-subgroup in the real part of the Jacobi variety
of this Riemann surface. It is shown that the KP-I flow with these initial data propagates almost
periodically.
Mathematics Subject Classifications (1991): Primary: 14H60; Secondary: 30F15, 58F07.
Key words: Kadomcev–Petviashvilli flow, Jacobi variety, infinite genus Riemann surfaces, Riemann–
Roch theorem.
1. Introduction
The Kadomcev–Petviashvilli partial differential equation (KP-I) is given by
ut = 14 uxxx + 32 ayy − 32 ux uxx , (1)
u = 2ax ; (2)
we impose the constraint that the mean value of u averaged in x-direction vanishes.
The KP-I equation reduces to the Korteweg–de Vries equation when there is no y-
dependence, ayy = 0. Bourgain [1] proved that the initial value problem with real
periodic L2 initial data u(x, y, t = 0) = q(x, y) is globally well posed. Feldman,
Knörrer and Trubowitz [2] showed that the KP flow propagates almost periodically
in t in case of real-valued real analytic periodic initial data. Here we prove almost
periodicity for real-valued periodic initial data q under the following regularity
condition: the Fourier coefficients q̂ of q have to satisfy k|j |r q̂(j )k1 < ∞ for
some r > 8; C 10 initial data fulfill this condition. The method differs from the
? Current address: Eurandom, PO Box 513, 5600 MB Eindhoven, The Netherlands. e-mail:
merkl@eurandom.tue.nl
246 FRANZ MERKL
one invented by Feldman, Knörrer and Trubowitz; it is based on an asymptotic

expansion of meromorphic functions over heat curves. The latter are Riemann
surfaces of infinite genus in general; they arise as spectral varieties of the heat
equation with a periodic external potential.
However, the basic idea can be seen already for Riemann surfaces of finite
genus; we sketch this classical situation as a motivation and a source of intuition
for the generalization to be treated:
Let X be a compact Riemann surface of genus g, P be a nonspecial divisor of
degree g on X and ∞ be a point on X (say with multiplicity 0 in D), 1/z be a
local coordinate near ∞, z(∞) = ∞. Then for x, y, t ∈ R sufficiently close to 0,
there is an unique meromorphic function ψ on X \ {∞} with pole orders bound by
z→∞
the divisor P such that ψ/et z +yz +xz −→ 1. The coefficient a = a(x, y, t) in the
3 2
Laurent expansion
ψ = exp(tz3 + yz2 + xz + az−1 + bz−2 + cz−3 + · · ·) (3)
at ∞ satisfies the KP-I-equation. More details are described in Section 8; a closely
related algebraic treatment of the KdV-equation is described in [14]; see [4 – 6, 8]
for other interesting approaches to solutions of the KP-equation and modified KP-
equation. In a geometric language, we have a one-point deformation at ∞ of the
line bundle O(P ) defined by the divisor P . The deformation is parametrized by
(x, y, t); the KP-I-flow is obtained by a deformation in t-direction. It corresponds
to a one-parameter-subgroup in the Jacobi variety of X.
We will generalize this to some Riemann surfaces of infinite genus: It is well
known (see, e.g., [2]) that the heat curves are invariant as the external potential
flows with the KP flow. Heat curves may be viewed as infinitely many handles
glued to a complex plane; the handles accumulate only at infinity, but they become
rather dense there. The meromorphic functions ψ that we are going to expand may
have one pole per handle; they grow like exp(tz3 + yz2 + xz) near infinity; z is a
coordinate in the complex plane with holes for the handles cut out. The presence
of infinitely many handles accumulating near infinity, each of them containing one
pole of ψ, is an obstruction for a convergent Laurent series (3), but we will derive
an asymptotic expansion of finite order:
ψ = exp(tz3 + yz2 + xz + az−1 + bz−2 + cz−3 + · · · + O(z−n )); (4)
it is valid outside the handles. The coefficient a = a(x, y, t) again solves the KP-
equation, and for an appropriate choice of the pole distribution we are able to fulfill
the given initial condition. The maximal order n of the expansion (4) that we can
control depends on the regularity of the potential.
We review some basic facts on heat curves and Bloch functions in the prepara-
tory Section 2, mostly without proofs. The main results are stated in Section 3. We
introduce line bundles with growth conditions at infinity in the first half of Section
4; this gives us a uniform geometric language to make the proofs more clear. The
main technical ingredient to derive the asymptotic expansion (4) is the Fredholm
AN ASYMPTOTIC EXPANSION FOR BLOCH FUNCTIONS 247
theory of the Cauchy–Riemann operator in certain line bundles over infinite genus
Riemann surfaces; especially we use a Riemann–Roch theorem for Riemann sur-
faces of infinite genus. This tool was prepared in [12]; we check the hypotheses
of the Riemann–Roch theorem for the relevant line bundles in the second half of
Section 4. We prove in Section 5 that meromorphic functions with the prescribed
pole distribution and with the prescribed asymptotics are unique; this is an infinite-
genus analogue for the nonspeciality of the divisor P in the above example. The
asymptotic expansion (4) and bounds for the derivatives of the coefficients with
respect to the parameters x, y, z, P are derived in Section 6, based on a Fredholm
theory for the Cauchy–Riemann operator. In Section 7, a L2 -bound for the error
term is used to derive a pointwise bound O(z−n ). Section 8 proves the KP-I equa-
tion for the Laurent coefficient a in the expansion. Compactness of the real part of
the Jacobi variety yields the almost periodicity of the KP-I flow; this is shown in
Section 9. Finally we prove that the initial conditions are fulfilled in Section 10.
Some parts of the material presented here are also contained in the author’s
doctoral thesis [13]; the latter was done under the supervision of Horst Knörrer at
the ETH Zürich.
2. Heat Curves
To start, we review the definition and some fundamental properties of heat curves.
For a detailed description of these constructions and for proofs we refer to [2], Part
III, §15, §16, [2], Part IV, §23, [7], Ch. I, §2, and [13].
Let 0 = γ1 Z ⊕ γ2 Z be a lattice in R2 , where γ1 = (γ11, 0), γ2 = (γ21, γ22 )
R γ11 γ11 , γ22 > 0. We fix a potential q ∈ ∗L (R ∗/ 0) with zero spatial mean,
2 2
with
0 q(x, y) dx = 0. For ξ = (ξ1 , ξ2 ) ∈ C × C let Fξ (q) denote the set of
all distributional solutions ψ in L2loc (R2 ) of the heat equation
(∂y − ∂x2 )ψ(x, y) + q(x, y)ψ(x, y) = 0 (5)
satisfying the Bloch conditions
ψ(x + γν ) = ξν ψ(x) (6)
for ν = 1, 2; x = (x, y). We define the heat curve:
H(q) := {ξ ∈ (C∗ )2 : Fξ (q) 6= {0}}.
The set H(q) is an analytic variety; it is invariant as q flows with respect to the
KP-flow; for proofs of these facts see [2], §15.
For free heat curves one has simply
2
H(0) = {(ξ1 , ξ2 ) = (eγ11 z , eγ21 z+γ22 z ) | z ∈ C} (7)
248 FRANZ MERKL
2
corresponding to the exponential solutions ψ(x, y) = exz+yz of the free heat
equation; there are countably many transversal self intersections of this surface
located at the points z = σ±j , j = (j1 , j2 ) ∈ (Z \ {0}) × Z with
iπ γ11 j2 γ21
σj := j1 + − ;
γ11 2γ22 j1 2γ22
at these special points the space Fξ (q) is two-dimensional; over all other points of
the free heat curve this space is one-dimensional.
The double point singularities of the free heat curve H(0) open up to handles
(or branch cuts) as the potential q is turned on; Krichever [7] has given an explicit
description of heat curves in terms of cutting and gluing operations:
Take a complex plane C (coordinate z) and cut it open along countably many
pairs of parallel lines [v1,j , v2,j ], [v1,−j , v2,−j ] with v1,j − v1,−j = v2,j − v2,−j ; the
index j runs over the index set J = N × Z. Then glue corresponding border lines
together by translation with v1,j − v1,−j to form a handle; we get an oval aj from
the the two identified branch cuts. In exceptional cases, v1,j = v2,j ; this means
that the two quadratic branch points degenerate to a double point singularity. For
the lines [v1,±j , v2,±j ] chosen appropriately, there is a natural biholomorphic map
between the Riemann surface obtained in this way and the heat curve H(q); let
[ [
z : H(q) \ aj → C \ [v1,±j , v2,±j ]
j j
denote this biholomorphic map outside the branch cuts. The coordinate z is charac-
terized modulo 2π i/γ11 by the condition ezγ11 = ξ1 ; here ξ1 is defined by the Bloch
condition (6). The additive constant in 2π iZ/γ11 is determined by the condition
|z(ξ )|→∞
ξ2 e−(γ21 z(ξ )+γ22 z(ξ ) )
2
−→ 1, (8)
ξ in the domain of z; the fact that one can fulfill (8) is a consequence of [2], Part
III, §16, Theorem 16.1. One should compare this with the special case (7).
All geometric facts on heat curves that we need are reviewed in Propositions
1 and 2 below. The first proposition describes the distribution of the ramification
points; it also provides bounds on the width |v1,j − v1,−j | of the handles:
PROPOSITION 1. Let q be real-valued and periodic, and P let r > 4. Assume that
the Fourier coefficients q̂(j ) of q satisfy q̂(0) = 0 and j |j |r |q̂(j )| < ∞. Then
|v1,j − v2,j | 6 O(|j1 |−1 |j |−r ) and |v1,j − σj | → 0 as |j | → ∞, j = (j1 , j2 ) ∈
(Z \ {0}) × Z. The branch cuts remain on lines parallel to the real axis:
iπj1
Im v1,j = Im v2,j = Im σj = .
γ11
Proof. See [2], Part IV, Theorem 23.4, Lemma 23.5 and [7], Ch. I, §2. 2
We remark that all handles are closed for q = 0: v1,j = v2,j = σj .

Roughly speaking, the proposition tells us that the heat curve H(q) asymptot-
ically resembles the unperturbed curve H(0) as |z| → ∞; the handles become
smaller but also more dense as |z| → ∞: asymptotically the potential term may be
treated as a small perturbation.
Complex conjugation ι: (ξ1 , ξ2 ) 7→ (ξ1 , ξ2 ) maps the heat curve H(q) onto
itself; it is described by z 7→ z in the coordinate z; it leaves the ovals aj lying over
the branch cuts pointwise fixed.
We introduce the domain U on which the asymptotic expansion will be valid;
we obtain it by removing neighbourhoods of the branch cuts: We choose some
small ε > 0; it will be held fixed. Then we remove all the ellipses described by
|z − vj,1 | + |z − vj,2 | < |vj,1 − vj,2 | + ε/|j1 | from the z-plane; for sufficiently small
ε these ellipses are pairwise disjoint, since neighbouring branch cuts to [vj,1 , vj,2 ]
have a distance of order const · |j1 |−1 . Let U denote the z-plane with all these
ellipses removed; let U0 be obtained by a similar construction with ε replaced by
ε/2; so smaller ellipses are removed here: U ⊆ U0 .
In the exceptional cases when some of the handles degenerate to double points,
we work with the normalization of H(q); no singularities of the Riemann surface
remain here.
The second proposition controls the Bloch solutions:
PROPOSITION 2. (1) When ξ ∈ H(q) is not a double point, the space of so-
lutions Fξ (q) has complex dimension 1. For 0 6= ψξ ∈ Fξ (q) the map Fx,y :
ξ 7→ ψξ (x, y)/ψξ (0, 0) is meromorphic on H(q) outside the double points; these
functions are regular at the normalization of the double points. There are at most
poles of first order at the zeros of ψξ (0, 0); there is precisely one zero of ψξ (0, 0)
over every cycle aj , and there are no other zeros.
(2) The Bloch solutions behave asymptotically over U like exponential solutions
of the free heat equation:
e−(xz(ξ )+yz(ξ ) ) ψξ (x, y)/ψξ (0, 0) = 1.
2
lim
z(ξ )→∞,ξ ∈U
We do not include a proof of this proposition in this article; but one may consult
[7], §2, especially Theorem 2.2 for a proof of the first statement in the case of
real analytic potentials; see also [13] for less regular potentials. The last citation
contains also a proof of part (2) of the proposition.
3. Statement of Results
We choose a divisor P of infinite degree over H(q) that assigns multiplicity 1 to
one point Pj on every oval aj , but 0 to all other points. The divisor of zeros of a
meromorphic function f is denoted by (f ); negative values mean poles.
We are going to prove:
THEOREM 1 (Asymptotic expansion of meromorphic functions). (1) Assume that
the potential q satisfies k|j |−r q̂(j )k1 < ∞ with r > 6. Let x = (x, y, t) ∈ R3 .
Then the complex vector space of global meromorphic functions ψ on the heat
curve with the following properties has the dimension 1:
250 FRANZ MERKL
• Poles of ψ are allowed only at the Pj : (ψ) > −P ;

• The growth rate of ψ outside the handles should be bounded by an exponen-
tial:
lim sup |ψ(z)e−(zx+z

2 y+z3 t )
| < ∞.
|z|→∞,z∈U
(2) This vector space is generated by a meromorphic function ψx,P with an

asymptotic expansion over U :
!
X
n−1
zx+z2 y+z3 t −l
ψx,P = e αl (x, P )z + rn (z, x, P ) (9)
l=0
and the with normalization condition α0 (x, P ) = 1. Let m = (m1 , m2 , m3 ) ∈ N30

be a multiindex; we set
∂ m1 ∂ m2 ∂ m3
D m :=
∂x m1 ∂y m2 ∂t m3
and v(m) := m1 + 2m2 + 3m3 . Then the error term rn satisfies for r > v(m) +
max{4, n + 3}:
lim sup |zn D mrn (z, x, P ))| < ∞.
|z|→∞,z∈U
(3) The derivatives D mak (x, P ) exist when v(m) < Q r − max{4, n + 1}; these
derivatives are continuous functions of (x, P ) ∈ R3 × j aj ; here the divisor P
is viewed as an element of the Cartesian product of all ovals aj endowed with the
product topology. The derivatives D m ψx,P (ξ ) exist when r > v(m) + 4Qtoo (with
poles for ξ at P ); they are are continuous functions of (x, P ) ∈ R3 × j aj and
meromorphic functions of ξ .
The next theorem shows us how to get a solution of the KP-I initial value
problem in terms of this expansion:
THEOREM 2 (Solution of the KP-I initial value problem). (1) For fixed P , the
first coefficient a(x) := α1 (x, P ) in the expansion (9) and u(x) := 2 ∂x
∂
a(x) yield
a solution of the KP-I equation (1). For fixed P , t and ξ , the function (x, y) 7→
ψx,P (ξ ) solves the heat equation
(∂y − ∂x2 )ψx,P (ξ ) + u(x, y, t)ψx,P (ξ ) = 0 (10)

and the Bloch conditions (6). In exceptional cases there is a pole at ξ ; then one
should substitute ψx,P by its residue
Qat ξ in any local coordinate.
(2) Let in the preceding P ∈ j aj denote the divisor of points ξ ∈ H(q)
where the Bloch solutions ψξ ∈ Fξ (q) vanish at the origin: ψξ (0, 0) = 0. Then the
constructed solution u of KP-I fulfils the initial condition u(x, y, 0) = q(x, y).
This description is helpful to obtain qualitative information on the KP-flow; we

describe the KP flow in terms of a Jacobian group operation over the real part of
the Jacobi variety of a heat curve; the latter is here represented by the Cartesian
product of the ovals aj .
THEOREMQ 3 (Jacobi group structure; almost periodicity of the KP flow). (1) Let
O, P , Q ∈ j aj . Then there is a nonvanishing global meromorphic function φ on
the heat curve which is bounded over U and satisfies (φ) > O −P −Q; it is unique
up to a multiplicative constant.Q Let P ⊕O Q := (φ)+P +Q−O Q denote
Q the divisor
Q
of zeros of φ; then P ⊕O Q ∈ j aj ; the operation Q ⊕O : j aj × j aj → j aj
defines a continuous Abelian group structure on j aj with neutral elementQ O.
(2) Let Q := (ψx,P ) + P denote the divisor of zeros of ψx,P . Then Q ∈ j aj ;
Q Q
we define the map v: R3 × j aj → j aj , v(x, P ) := Q. Then v is a continuous
Q
group operation of (R3 , +) on j aj : the relations v(x + y, P ) = v(x, v(y, P ))
and v(x, P ) = v(x, O) ⊕O P hold.
(3) Let u denote the solution of the KP-I initial value problem with real periodic
initial data q with k|j |−r q̂(j )k1 < ∞, r > 8. Then we have almost periodicity of
the solution in the following sense: For every ε > 0 there is L > 0 so that every
real interval of length L contains a number τ with
sup |D mu(x, y, t + τ ) − D m u(x, y, t)| < ε
(x,y,t )∈R3
for all multiindices m ∈ N30 with v(m) < r − 5.
4. Line Bundles with Growth Conditions

We get a uniform description of the various growth conditions of meromorphic
functions near infinity over U0 when using “bounded line bundles”. We define a
bounded line bundle to be a pair (L, φ), where L is a holomorphic line bundle over
the heat curve and φ is a basis section defined near infinity over U0 ; two such pairs
(L, φ), (L0 , φ 0 ) are called isomorphic when there is a line bundle isomorphism f :
L → L0 so that (f∗ φ)/φ 0 and (f ∗ φ 0 )/φ are bounded. By definition, global sections
in (L, φ) are global sections ψ in L such that ψ/φ is bounded over U0 near infinity.
Four examples are fundamental for us:
(1) For x = (x, y, t) ∈ R3 , let Ex = (O, exz+yz +t z ); here O denotes the trivial
2 3
holomorphic line bundle.

(2) Let P be a divisor that is supported on the union of the ovals aj . Although P
may have infinite degree, its support should only have finitely many points per
handle. We define Ob (P ) := (O(P ), 1); by definition, holomorphic sections
in O(P ) are meromorphic functions ψ over the heat curve with (ψ) > −P .
(3) Ob1,0 := (O 1,0, d 1z ); O 1,0 denotes the line bundle of holomorphic one forms.
(4) Ob (n∞) := (O, zn), n ∈ Z; we view n∞ as a divisor supported at infinity.
252 FRANZ MERKL
From these basic examples we build the bounded line bundles of interest by
tensor products; the tensor product for two pairs is defined in a natural way by
(L, φ) ⊗ (L0 , φ 0 ) = (L ⊗ L0 , φ · φ 0 ), and the dual of (L, φ) is defined by (Ľ, 1/φ).
We abbreviate L(P ) := L ⊗ Ob (P ) and L1,0 := L ⊗ Ob1,0 for every bounded line
bundle L; the space of global holomorphic sections in L is denoted by 0(L).
The first statement in Theorem 1 claims that dim 0(Ex (P )) = 1; the Riemann–
Roch theorem for Riemann surfaces of infinite genus proved in [12] yields
dim 0(Ex (P )) − dim 0(E−x
1,0
(−P )) = 1;
one should note that every handle receives precisely one point of the divisor P ; in
the language of reference [12] this implies that the codegree of P vanishes; this
justifies the right-hand side 1 in the Riemann–Roch formula. Similarly the first
part of Theorem 3 states dim 0(Ob (P + Q − O)) = 1; the Riemann–Roch theorem
yields in this case:
dim 0(Ob (P + Q − O)) − dim 0(Ob1,0(−P − Q + O)) = 1;
one should note that the codegree of P + Q − O vanishes too, since the number of
points per handle of the divisor counted with multiplicity is 1.
It remains to check the hypotheses of the Riemann–Roch theorem for Ex (P )
and Ob (P + Q − O). This is a rather technical matter; the reader who searches
only an overview may want to skip the rest of this section except the introduction
of the handle coordinates zj : Uj → C in Equations (11); only Lemma 1 will be
used later too.
We show that the hypotheses (X1)–(X6) in [12] are fulfilled: First, coordinates
zj in the handles are required: We introduce scaling constants cj = c−j > 0 for
every handle; they should satisfy 0 < lim infj cj /|j1 | 6 lim supj cj /|j1 | < ∞;
this means that the distance of neighbouring handles is of order cj−1 in the z-plane.
More details for cj are specified below. Set sj := (v1,j + v2,j )/2 (center of the j -th
√
branch cut) and tj := |v2,j − v1,j |cj /2 (rescaled width of the j -th branch cut);
define zj by the equations
tj p
(z − sj )cj = zj + for |zj | > tj ,
zj
tj p
(z − s−j )cj = zj + for |zj | 6 tj (11)
zj
in a neighbourhood of the oval aj . We have the equation z−j zj = tj ; the ovals aj
√
are described by |zj | = tj . The equations |zj | = const describe ellipses with
focii at the ramification points; for example the boundary ellipses of U0 are given
by |zj | = ε1 (j ) for some constants ε1 (j ) = ε1 (−j ).
Define the j -th handle Uj to consist of all ξ ∈ H(q) such that |zj (ξ )| < 1 and
|z−j (ξ )| < 1; we restrict the coordinate zj to be defined only in this domain. We
choose the scaling constants cj so that:
• the constants ε1 (j ) do not depend on j ∈ J at least for large |j |;
• ε1 (j ) < 1;
• the handles Uj are pairwise disjoint and disjoint from U .
This all can be done if the constant ε in the definition of U0 is sufficiently small.
The intersection Uj ∩ U0 consists of two connected components Vj , V−j ; these are
the annuli V±j = {ξ ∈ Uj : 1 > |z±j (ξ )| > ε1 (j )}. Obviously we can include the
elliptic annuli z[Vj ] – at least for large |j | – in pairwise disjoint discs in the z-plane
centered at sj with a radius rj 6 O(|j1 |−1 ).
These considerations show that the hypotheses (X1)–(X3) in [12] are fulfilled.
We check (X4): We look for bounds on the derivatives of the coordinate changes:

dz tj
sup cj − 1 = sup 2 6 const · tj
Vj dzj V j zj
and

d2 z tj
sup cj 2 = 2 sup 3 6 const · tj .
Vj dzj V j zj
√
Using Proposition 1 and the fact that |cj | scales like |j1 |, we get tj 6 O(oj ) for
large |j | with oj := |j |−r ; but also tj 6 O(oj ) for large |j |; from this we obtain
the hypothesis (X4) in [12].
The hypothesis (X5) in [12] states parabolicity of the surface; thisR means: for all
N b X and ε > 0 there is χ ∈ Cc∞(X, [0, 1]) with χ|N = 1 and X dχ ∧ ∗dχ <
ε. This is easy to prove by explicitly constructing χ from a cutoff-function in C
interpolated through the handles; we refer the reader either to [2] or to [13].
All quantitative estimates on the scaling constants that are required for the
Riemann–Roch theorem are assembled in the hypothesis (X6) in [12]; we check
them here:
• αj := |cj ||sj |2 is bounded from below by a positive constant since both factors
are bounded from below;
• tj 6 O(oj2 );
P have |α2j | 6P O(|j1 ||s
• We j | ) 6 O(|j ||σj | ) 6 O(|j | ); consequently
2 2 3
j ∈J |j |
6−2r
j ∈J (αj oj ) < ; the sum runs over a two-dimensional lattice;
hence it is convergent for r > 4.
• oj αj2 6 O(|j |6−r ) is bounded for r > 6;
• lim supj |rj /sj | < 1, since |rj | → 0, while |sj | → ∞.
Q
Let P , Q, O ∈ j aj . The bounded line bundles Ob (P ) and Ob (P + Q − O) were
already treated as examples in the cited reference based on the hypotheses (X1)–
(X6) only; but it remains to check the remaining hypotheses (L1)–(L2) in [12] of
the Riemann–Roch theorem for the “deformed” bounded line bundles Ex (P ): we
observe that deformations with asymptotically almost constant transition functions
preserve the hypotheses of the Riemann–Roch theorem: let the bounded line bundle
(L, ψ0 ) satisfy the hypotheses of the Riemann–Roch theorem with basis sections
254 FRANZ MERKL
ψj over Uj , and let (E, φ0 ) be a bounded line bundle with basis sections φj over Uj
that satisfy the following bounds: for some constants ζj of modulus 1 and ν = ±1,
ν
φ0 d φ0ν
−1 ν
sup oj sup ν − ζj < ∞ and sup oj sup −1
ν < ∞. (12)
j Vj φj j Vj dzj φj
Then (L ⊗ E, ψ0 · φ0 ) satisfies the hypotheses of the Riemann–Roch theorem too,

using the basis sections ψj ·φj over Uj . One should also notice that the codegree of
the bundles is unchanged by this deformation with (E, φ0 ) whenever the transition
functions φ0 /φj have a holomorphic logarithm defined over Vj ∪ V−j .
The following lemma is the key to check the estimates (12) for the bounded line
bundle Ex ; but it will also play an essential role later:
LEMMA 1. (1) There exist imaginary numbers αj,l = −α−j,l and holomorphic
functions ηj,l = η−j,l : Uj → C for l ∈ Z with |αj,l | 6 |sjl |,
sup |zl − ηj,l − αj,l | 6 O(|sjl−1 tj /cj |),

Vj
and (13)
d l

sup (z − ηj,l ) 6 O(|sjl−1 tj /cj |).
Vj dzj
When l 6 r + 1 this implies

sup oj−1 sup |zl − ηj,l − αj,l | < ∞
j Vj
and (14)
d l
−1
sup oj sup (z − ηj,l ) < ∞.
j dz
Vj j
The functions ηj,l can be chosen antisymmetric with respect to complex conjuga-
tion: ηj,l (ι(ξ )) = ηj,l (ξ ), ξ ∈ Uj . (2) Similarly there exist imaginary numbers
αj,log = −α−j,log and holomorphic functions ηj,log = η−j,log : Uj → C with
sup oj−1 sup | Log z − ηj,log − αj,log | < ∞
j Vj
and (15)
d
sup oj−1 sup (Log z − ηj,log ) < ∞,
j Vj dzj
and |αj,log | 6 π . Here “Log” means the principal branch of the logarithm. The
functions ηj,log can be chosen antisymmetric with respect to complex conjugation,
too.
Proof. (1) We define ηj,l := wjl + w−j l
− Re s−j
l
with wj := zj /cj + sj over Uj ,
and αj,l := −i Im s−j ; obviously |αj,l | 6 |s−j | = |sjl | holds. We get supVj |wj −
l l
z| 6 O(|tj /cj |) and supVj | dzdj (wj − z)| 6 O(|tj /cj |) since wj − z = −tj /(zj cj )
over Vj ; recall infj infVj |zj | > 0.
This implies
Z 1
sup |wjl − z | 6 |l| sup |wj − z|
l
|(1 − t)z + twj |l−1 dt 6 O(|sjl−1 tj /cj |);
Vj Vj 0
one should note that |sj |−1 |(1 − t)z + twj | is bounded from above and from below
on Vj uniformly in j by positive constants.
The derivative is estimated by

d l−1 d
l−1 d

sup l
(wj − z ) 6 |l| sup z
l
(wj − z) + |wj − z |
l−1
wj
Vj dz j Vj dz j dz j
6 O(|sjl−1 tj /cj |) + O(|sjl−2 tj /cj2 |) 6 O(|sjl−1 tj /cj |);

we used infj |sj |, infj |cj | > 0.
Similarly w−j − s−j = z−j /cj = tj /(zj cj ) implies supVj |w−j − s−j | 6
O(|tj /cj |) and supVj |d/dzj w−j | 6 O(|tj /cj |); therefore we conclude supVj |w−j
l
−
s−j | 6 O(|s−j tj /cj |) = O(|sj tj /cj |) and supVj |d/dzj w−j | 6 O(|sj tj /cj |);
l l−1 l−1 l l−1
we utilized that |w−j |/|sj | is bounded from both sides on Vj uniformly in j .

Adding these estimates yields
sup |zl − ηj,l − αj,l | = sup |(zl − wjl ) − (w−j
l
− s−j
l
)| 6 O(|sjl−1 tj /cj |)
Vj Vj
and

d l d l d
sup (z − ηj,l ) = sup (z − wj ) −
l
w−j 6 O(|sjl−1 tj /cj |);
l
Vj dzj Vj dzj dzj
these bounds hold for all l ∈ Z. The estimates (14) follow from O(|sjl−1 tj /cj |) 6
O(|j |−2r+l−1) 6 O(|j |−r ) = O(oj ) with the help of |sj | 6 O(|j |), |cj−1 | 6 const,
tj 6 O(|j |−2r ), and l − 1 6 r.
The symmetry properties for ηj,l follow from w−j (ι(ξ )) = wj (ξ ) and s−j = sj .
(2) The proof of these bounds is similar to the previous one; one just has to
replace l-th powers in the proof by logarithms and the (l − 1)-st powers by recip-
rocals. One should note that z[Vj ] does not meet a neighbourhood of the negative
real axis; so Log z is continuous and bounded on every fixed Vj . 2
We introduce the basis sections φ0 := exz+yz +t z over U0 and φj :=

2 3
xηj,1 +yηj,2 +t ηj,3

e over Uj in the bounded line bundle Ex , x = (x, y, t) ∈ R3 ; the
lemma immediately yields the bounds (12) with ζj = exαj,1 +yαj,2 +t αj,3 ; we need
r > 2 here.
5. Nonspeciality of Divisors
We are going to prove uniqueness of holomorphic sections in Ob (P ), Ex (P ) and
Ob (P + Q − O), starting with the simplest case:
256 FRANZ MERKL
LEMMA Q 2. 0(Ob (P )) consists only of the constant functions for every divisor
P ∈ j aj .
Proof. We have to show dim 0(Ob (P )) = 1. By the Riemann–Roch theorem
dim 0(Ob (P )) − dim 0(Ob1,0(−P )) = 1; so it suffices show 0(Ob1,0(−P )) = {0}.
So assume that there was an ω ∈ 0(Ob1,0 (−P )), ω 6= 0. We may assume that ω
is antisymmetric, ι∗ ω = −ω, by replacing it either by ω − ι∗ ω or by iω − ι∗ (iω); at
least one of these two R forms is nonzero. Since all aj -cycles are symmetric, ι∗ aj =
aj , the aj -periods aj ω are imaginary. We cut the heat curve open along all ovals
aj ; the resulting surface Y may be viewed as the complex plane with all branch cuts
[v1,j , v2,j ] cut open. The homology of Y is generated by these ovals; and σ := Re ω
has zero aj -periods; so σ is exact over Y , say σ = df on Y . ω ∈ 0(Ob1,0(−P ))
implies |ω/(z−2 dz)| 6 const over U0 ; by integrating this bound over semi-infinite
pathes γ in U0 we get supU0 |zf | < ∞; we have fixed an integration constant. To
derive this bound, one cannot use radial semiinfinite lines and arcs centered at 0
in the z-plane for γ in general; these might hit the elliptic holes in U0 . But one
may modify the lines making a little detour around the elliptic holes; this locally
increases the length of the path only at most by a multiplicative constant.
Using the maximum principle for bounded subharmonic functions on parabolic
surfaces we conclude supU0 f = sup∂U0 f ; the boundary ∂U0 consists of the union
of all the ellipses ∂(Uj \ U0 ).
The sup∂U0 f is reached at one point ξ ∗ ∈ ∂(Uj ∗ \ U0 ) for some handle index
∗
j , since otherwise we would have sup∂U0 f = lim sup|j |→∞ sup∂(Uj \U0 ) f = 0;
the latter is impossible since f does not vanish everywhere and is antisymmetric,
ι∗ f = −f .
The symmetric oval aj ∗ divides the handle Uj ∗ into two pieces; let W be the
half of Uj ∗ in which ξ ∗ lies. f |W has to reach its maximum at a point ζ ∗ on
the boundary ∂W . ∂W consists of the symmetric oval aj ∗ and of one of the two
connected components of ∂Uj ∗ . f |W cannot reach its maximum at a ζ ∗ ∈ ∂Uj ∗ ,
since this would imply f (ζ ∗ ) > f (ξ ∗ ) = sup∂U0 f, which is impossible as ζ ∗ is
an interior point of U0 and f is nonconstant. Hence, ζ ∗ ∈ aj ∗ ; to be precise, aj ∗
is doubled, consisting of two opened branch cuts, since we have cut X open along
the aj ; we consider only the branch cut in the border of W .
f is constant along aj ∗ : to show this, take any vector v tangent to aj ∗ . Then
ι∗ v = v, hence hdf, vi = hι∗ df, ι∗ vi = −hdf, vi, i.e. hdf, vi = 0.
We know that f |W reaches its maximum on the whole oval aj ∗ , but does not
reach it on W \aj ∗ . We also know that df = σ = Re ω has a zero at one point Pj ∗ ∈
aj ∗ of the divisor P . The following lemma, applied to f written in logarithmic polar
√
coordinates z±j ∗ = tj ∗ el+iϕ , shows that this is impossible; this contradiction
yields that ω does not exist. 2
LEMMA 3 (Nonvanishing derivatives at the symmetric ovals). Let u be a noncon-

stant 2π i-periodic real harmonic function on [0, L] + iR with existing derivatives
at the boundary. Assume that u(z) > 0 for Re z > 0 and u(z) = 0 for Re z = 0.
Then

∂
∀ϕ ∈ R : u(l + iϕ) > 0.
∂l l=0
Proof. By the maximum principle we get u(z) > 0 for 0 < Re z < L, and
therefore for every fixed 0 < L1 < L:
ε := inf u(z) > 0.
Re z=L1
Let us compare the following two Dirichlet boundary conditions on the compact
cylinder ([0, L1 ] + iR)/2π iZ:

0, Re z = 0,
r1 (z) =
ε, Re z = L1 ,
r2 (z) = u(z), Re z ∈ {0, L1 },
r1 6 r2 implies that the (unique) solutions u1 resp. u2 of the Dirichlet problem
on the cylinder with boundary conditions r1 resp. r2 satisfy u1 6 u2 . But u1 (z) =
ε
L1
Re z, u2 (z) = u(z), hence
∂u u2 (l + iϕ) u1 (l + iϕ) ε
(l + iϕ) = lim > lim = > 0. 2
∂l l&0 l l&0 l L1
To show uniqueness of global sections in the more complicated bounded line

bundles of interest, we reduce them to the special case Ob (P ):
Q
PROPOSITION 3. Let L equal Ex (P ) or Ob (PQ+Q−O) where P , Q, O ∈ j aj .
Q to Ob (R) for some R ∈ j aj . In particular, dim 0(L) = 1.
Then L is isomorphic
The divisor R ∈ j aj is uniquely determined.
Proof. By the Riemann–Roch theorem, dim 0(L) = 1 + dim 0(Ľ1,0) > 1. Let
ψ ∈ 0(L) \ {0}. We view ψ as a meromorphic function over the heat curve. ψ ◦ ι
is another section in 0(L); this is due to complex conjugation symmetry: ι∗ P = P ,
ι∗ Q = Q, ι∗ O = O; x is real. We may assume reflection symmetry ψ = ψ ◦ ι
by replacing ψ by ψ + ψ ◦ ι or iψ + iψ ◦ ι; at least one of these two functions is
nonzero. Q
We are going to show that ψ has zeros at a divisor R ∈ aj ; this means that
ψ ∈ 0(L(−R)): The ovals aj are pointwise fixed under complex conjugation ι;
hence ψ is real-valued (with poles) over aj . We examine the real analytic map
ψ|aj → R ∪ {∞}; we view R ∪ {∞} as the one-point-compactification of R:
The degree ν of the divisor (ψ|aj ) must be even; by definition this is the number
of zeros of ψ on aj minus the number of poles of ψ on aj (counted with multiplic-
ity): the parity of ν equals the parity of the number of sign changes of ψ|aj , which
is an even number.
258 FRANZ MERKL
Let us examine the case L = Ex (P ) first: The restriction of the divisor P to aj

consists of only one single point Pj ; since ν is even, (ψ|aj ) cannot equal Pj ; we
must have (ψ|aj ) > Rj − Pj for some Rj ∈ aj since (ψ|aj ) > −Pj ; here “>”
denotes the partial order relation for divisors.
The case L = Ob (P +Q−R) is treated similarly: The restriction Pj +Qj −Oj
of the divisor P + Q − O to aj has odd degree; we conclude that (ψ|aj ) > Rj −
(Pj + Qj − Oj ) for some Rj ∈ aj . Q
We conclude that ψ ∈ 0(L(−R)) with R = (Rj )j ∈ j aj .
Set F := Ľ(R + P ), i.e. F = E−x (R) in the case L = Ex (P ); but F =
Ob (R + O − Q) in the case L = Ob (P + Q − O). We get again dim 0(F) > 1
by the Riemann–Roch theorem; let φ ∈ 0(F) \ {0} with φ ◦ ι = φ. One can repeat
the above considerations withQψ and L replaced by φ and F: we conclude that
φ ∈ 0(F(−S)) for some S ∈ j aj .
Then ψ · φ ∈ 0(L(−R) ⊗ F(−S)) = 0(Ob (P − S)) ⊆ 0(Ob (P )) = C; see
Lemma 2. This implies that ψ · φ is a constant; it cannot be zero since ψ 6= 0 and
φ 6= 0. Without loss of generality, let us assume ψ · φ = 1. 1 ∈ 0(Ob (P − S))
leads to P > S; therefore P = S, since both divisors P and S have precisely one
point per handle. The multiplication with ψ ∈ 0(L(−R)) defines a line bundle
homomorphism ψ: Ob (R) → L; it respects the growth condition at infinity. Simi-
larly the multiplication with φ ∈ 0(F(−P )) defines a line bundle homomorphism
φ : L → L ⊗ F(−P ) = Ob (R) respecting the growth conditions; these two homo-
morphisms must be isomorphisms since φ ◦ ψ = idOb (R) and ψ ◦ φ = idL . Thus
Ob (R) and L are isomorphic; Lemma 2 yields dimQ 0(L) = dim 0(Ob (R)) = 1.
It remains to show that the properties R ∈ ∼
j aj and L = Ob (R) deter-
Q
mine R uniquely: Assume L ∼ = Ob (R̃) for another R̃ ∈ j aj . We conclude
∼
Ob (R) = Ob (R̃). Any isomorphism between these two bounded line bundles is
a multiplication with a χ ∈ 0(Ob (R̃ − R)) ⊆ 0(Ob (R̃)) = C; so 0 = (χ) =
R̃ − R. 2
The proof of the proposition yields some statements in Theorem 3:

In the case L = Ob (P + Q − O) we know (φ) > P + Q − O − R and
−(φ) = (ψ) > O + R − P − Q; thus (ψ) = O + R − P − Q. Therefore
the Jacobi group operation P ⊕O Q := R is well defined; the inverse of P is
defined to be O P := (χ) + 2O − P , where χ ∈ 0(Ob (2O − P )) \ {0}; it is
easy to check the group axioms using that P ⊕O Q Q is uniquely characterized by
∼
Ob (P + Q − O) = Ob (P ⊕O Q) and P ⊕O Q ∈ j aj .
We examine case L = Ex (P ): The inequalities (φ) > P − R and −(φ) =
(ψ) > R − P imply (ψ) = R − P . The operation v in Theorem 3 is defined by
v(x, P ) := R; it is uniquely characterized by Ex (P ) ∼ = Ob (v(x, P )). We check
that this is a group operation: Ob (v(x + y, P )) = Ex+y (P ) ∼
∼ = Ex ⊗ Ey (P ) ∼=
∼ ∼
Ex (v(y, P )) = Ob (v(x, v(y, P ))) and Ob (v(x, O) ⊕O P ) = Ob (v(x, O) + P −
O) ∼= Ex (O) ⊗ Ob (P − O) ∼ = Ex (P ) ∼
= Ob (v(x, P )).
6. Asymptotic Expansion
We want to determine the asymptotic behaviourQ of ψ ∈ 0(Ob (P + Q − S)) or ψ ∈
0(Ex (P )) for varying x ∈ R3 , P , Q, S ∈ j aj . Both cases are included when
P
we consider line bundles F = Ex ( N nκ P κ ) with divisors P κ = (Pjκ )j ∈J ∈
Q PN κ=1
j aj and multiplicities nκ ∈ Z, κ=1 nκ = 1. We are interested in an asymptotic
expansion of the global section in F (which is unique up to a constant) and its
dependence on the parameters x and P κ . We avoid to examine even more general
bounded line bundles to keep the treatment
Q simpler.
We fix a reference divisor O ∈ j aj ; the calculations will be done using the
fixed bounded line bundle Ob (O)⊗Ob (−n∞), n ∈ N; this is convenient to analyze
the parameter-dependence.
We will define only smooth global sections ϕ in F, ϕl in F(−l∞) and µj
in F(O); these sections will be “almost holomorphic” sections in F; with these
functions we build an ansatz for the global holomorphic section ψ in F:
X X
n−1
ψ = rϕ + βj µj + αl ϕl . (16)
j ∈J l=0
The complex coefficients αl and βj and the error term r = rn have to be deter-
mined; r is an only smooth section in Ob (O) ⊗ Ob (−n∞).
We have dropped the parameters x, P for simplicity of the notation; but we will
write them explicitly at some places to avoid ambiguity. Detailed specifications of
the ansatz functions µj , ϕl , ϕ are given below; but we describe the leading idea
first:
• The first summand rϕ in (16) will describe the error term in the expansion
2 3
(9). ϕ equals ezx+z y+z t over the domain U where the asymptotic expan-
sion is valid, but it is interpolated through the handles only smoothly but
“almost P meromorphically” with zeros/poles with orders bounded by the di-
visor − N κ
κ=1 nκ P . The factor r in the error term should satisfy a bound
−n
O(|z |) over U ; it will be only smooth in the handles with poles allowed at
the reference divisor O. The finiteness of a certain weighted Sobolev-norm
will guarantee these conditions on the error term r. We allow an additional
zero per handle of ϕ (below described by the divisor P 0 ) to get equal numbers
of zeros and poles per handle (counted with multiplicity) in the ansatz func-
tion; theP location of this zero is chosen in a way that the bounded line bundle
E(x,y,t )( N κ
κ=0 nκ P ) is almost isomorphic to Ob , meaning that it has a almost
holomorphic global basis section ϕ.
• The lth term in the last sum in (16) will describe the lth term in the ex-
pansion (9). To get this, we will choose the ansatz functions ϕl to equal
z−l ezx+z y+z t near infinity over U ; they are interpolatedPthrough the handles
2 3
with zeros/poles of orders bounded by the divisor − N κ

κ=1 nκ P , too; the
construction of the ϕl is similar to the construction of ϕ.
260 FRANZ MERKL
P
• The remaining summand j ∈J βj µj in the ansatz (16) will vanish over the
domain U of the asymptotic expansion; the ansatz functions µj will be sup-
ported in the j th handle with singularities at the Pjκ and Oj ; these ansatz
functions µj are only needed as counterterms to remove possible poles at O.
These functions are chosen to be “almost meromorphic” too.
It is necessary to allow one pole per handle in the error term r, since we want the
Cauchy–Riemann operator in a corresponding space of sections to be a Fredholm
operator: We know this to be the case for Sobolev spaces over Ob (O)⊗Ob (−n∞),
but not for Sobolev spaces over Ob (−n∞). S
The ϕl and µj should be meromorphic outside j (U0 ∩ Uj ). We will have to
solve a linear system for the unknown complex coefficients αl , βj and the unknown
error term r; the conditions to fulfill are the following:
• the Cauchy–Riemann equation ∂ψ = 0 (to be interpreted in an appropriate

space of sections);
• the condition resOj ψ = 0 that there should be no pole at Oj ; the residue has
to be interpreted for sections in F.
This leads to the following linear system, written in block matrix form:
 
r
Mx,P  (βj )j ∈J  = 0 (17)
(αl )l=0,...,n−1
with the matrix

M11 M12 M13
Mx,P =
M21 1 0

∂ + ϕ −1 ∂ϕ (ϕ −1 ∂µj )j ∈J (ϕ −1 ∂ϕl )l=0,...,n−1
= .
((resOj µj )−1 ϕ(Oj ) resOj )j ∈J 1 0
We will interpret this matrix as a Fredholm operator M: V ⊕ `2w (J ) ⊕ Cn →

B̌ ⊕ `2w (J ) with appropriate Hilbert spaces `2w (J ), V and B̌. (The index “w” stands
for “weighted”.)
CHOICE OF THE HILBERT SPACES
of J -indexed sequences β =
The space `2w (J ) is defined to be the Hilbert space P
(βj )j ∈J ∈ CJ with the weighted norm kβk22,w := j ∈J |cj−2 sj2n−4 | · |βj |2 being
finite.
We define the smooth line bundle L := O(O) and L0,1 over the heat curve; the
latter denotes the bundle of L-valued one-forms of type (0, 1). It was shown in [12]
that there are Hermitian metrics | · |A on L, | · |B̌ on L0,1 and a volume form with
the following properties:
• Let B̌ be Rthe completion of Cc∞ (H(q), L0,1 ) with respect to the norms
kωk2B̌ := H(q) |ω|2B̌ . Further let V be be the completion of Cc∞ (H(q), L)
R
with respect to the norm krk2V := H(q) (|r|2A + |∂r|2B̌ ). Then the Cauchy–
Riemann operator ∂: V → B̌ is a Fredholm operator of index 1 − n.
• ∼ d2 (1/z) over U0 near infinity, ∼ |cj−2 sj−4 | d2 zj over Vj ; we use the
abbreviation d2 w := d Re w∧d Im w; the relation “∼” means that the quotient
of both sides is bounded from above and from below by positive constants not
depending on the handle index j .
• |1|A ∼ |zn| over U0 near infinity, |1|A ∼ |sjn | over Vj ;
• |dz|B̌ ∼ |zn+2 | over U0 near infinity, |dzj |B̌ ∼ |cj sjn+2 | over Vj ;
DEFINITION OF THE ANSATZ FUNCTIONS

P
We choose four different smooth partitions of unity 1 = χ0,k + j ∈J χj,k (k =
0, 1, 2, 3) adapted to the open covering U0 , Uj (j ∈ J ); they should be symmetric
with respect to complex conjugation: χj,k = χ−j,k = χj,k ◦ ι (j ∈ J ). The support
of χj,k+1 and of χ0,k (k = 0, 1, 2) should be disjoint. The derivatives |dχj,k /dzj |
should be bounded on Vj uniformly for j ∈ J ; such a choice is possible since
zj [Vj ] contains a fixed annulus centered at 0 at least for large |j |; one may choose
χj,k over Vj to depend on |zj | and k only.
We define ϕ := φ0 and ϕl := χ · φl with functions φl that equal z−l exz+yz +t z
2 3
over U but are interpolated almost meromorphically through the handles; the smooth
cutoff function χ serves only to remove the singularity of z−l near z = 0; so
the function χ should vanish near z = 0; but it should equal 1 outside a small
neighbourhood of z = 0. The functions φl will be specified below.
We abbreviate:
x̂ = (−l, x) := (−l, x, y, t);
ẑ = (Log z, z) := (Log z, z, z2, z3 );
η̂ = (ηj,log , η) := (ηj,log , ηj,1 , ηj,2 , ηj,3 );
α̂ = (αj,log , α) := (αj,log , αj,1 , αj,2 , αj,3 );
ûv̂ = ulog vlog + uv := ulog vlog + u1 v1 + u2 v2 + u3 v3
for û = (ulog , u) = (ulog , u1 , u2 , u3 ), v̂ = (vlog , v) = (vlog , v1 , v2 , v3 ); û, v̂ ∈
C4 ; u, v ∈ C3 ; the data ηj,... and αj,... were
Q introduced in Lemma 1. We define
an additional divisor P 0 = (Pj0 )j ∈J ∈ j aj with multiplicity n0 = −1 by the
relation
YN
zj (Pjκ )nκ = e2x̂ α̂ ; (18)
κ=0
262 FRANZ MERKL
one should note that P 0 depends on x and on l, although we suppress writing this
explicitly. The divisor P 0 serves as a location for an additional zero per handle of
the ansatz functions.
Morally we define:
X
φl = exp χ0,1 log φ0,l + χj,1 log φj,l
j
with
φ0,l := ex̂ ẑ = z−l exz+yz +t z ,

2 3
(19)
Y
N
φj,l := ex̂ η̂+x̂ α̂
· (zj − zj (Pjκ ))−nκ
κ=0
Y
N
= ex̂ η̂−x̂ α̂ · (z−j − z−j (Pjκ ))−nκ ; (20)
κ=0
to see the last equation one uses zj − zj (Pjκ ) = −zj · zj (Pjκ ) · (z−j − z−j (Pjκ ))/tj ,
P
the relation (18), and N κ=0 nκ = 0. To make φj,l well-defined, we have to specify
which branch of the logarithm should be taken over Vj ; so we set over Vj :
X XN !
zj (Pjκ )
φl := exp χ0,1 · x̂ ẑ + χj,1 · x̂ α̂ + x̂ η̂ − nκ Log 1 − .
j κ=0
zj
“Log” means the principal branch of the logarithm; one should note that |zj | >
√
|zj (Pjκ )| = tj over Vj .
φl is meromorphic (poles at P ) outside the union of all Vj . We calculate to
which extent it deviates from being holomorphic over Vj ;

−1 ∂φl XN
zj (Pjκ )
φ = ∂χj,1 · − x̂ ẑ + x̂ α̂ + x̂ η̂ − nκ Log 1 −
l ∂z ∂z z
j j j
κ=0 √
tj
6 const1 · |x̂(−ẑ + α̂ + η̂)| + log 1 −
|zj |
6 const · |j |−r · (|x| + 1); (21)
we used Lemma 1; recall that |∂χj,1 /∂zj | is bounded.
BOUNDS FOR THE MATRIX ELEMENTS AND THEIR DERIVATIVES
We need to estimate also derivatives with respect to x and Pjκ over Vj ; this is a
rather technical matter. The reader might want to read only the statement of the
four lemmas in this section, but skip their proofs.
Let m = (m1 , m2 , m3 ) ∈ N30 be a multiindex. We write parametrize the points

√
Pjκby zj (Pjκ ) =: tj eipj,κ , pj,κ ∈ R. Let ν = (ν1 , . . . , νN ) ∈ NN
0 be another
multiindex; we introduce the abbreviation:
ν
YN
∂ νκ
∂j := νκ .
κ=1
∂pj,κ
One should notice that there is no derivative with respect to pj,0 ; this variable is
not considered independent, since
X
N
pj,0 = 2i x̂ α̂ + nκ pj,κ (22)
κ=1
(modulo 2π ), see Equation (18). We examine the case (ν, m) 6= (0, 0):

ν m −1 ∂φl ∂χj,1 ν m
∂ D φ =
j l
∂zj ∂z · (∂j D x)(−z + α + η) −
j

XN
ν m zj (Pjκ )
− nκ ∂j D Log 1 − .
zj
κ=0
ν
∂j D m x equals 0 unless m1 +m2 +m3 = 1 and ν = 0; Lemma 1 yields the following
bound:
sup |(∂j D m x)(−z + α + η)| 6 const · |j |−r · δν,0 .
ν
Vj
To estimate the derivatives of the Log-terms over Vj , set f (t) := Log(1 − t) and
√ ν
gκ := tj eipj,κ /zj . By the chain rule, ∂j D m f (gκ ) is a finite sum of terms of the
form
Y
k
ν
f (k) (gκ ) ∂j ` D m` gκ
`=1
Pk P
with k > 1, `=1 ν ` = ν and k`=1 m` = m. The |gκ | are bounded over Vj
√
uniformly in j by a constant less than 1; one should note |zj | > const > tj over
Vj . We get a bound |f (k) (gκ )| < constk uniformly in j over Vj .
Let us estimate in the case κ = 0:
ν
∂j ` D m` g0 = (−2α)m (in)ν ` g0 ;
k
Y
ν ` m`
∂j D g0 = |nν (−2α̂)m g0k | 6 constν,k,m · |sj |v(m) |j |−r

`=1
Q
we used the notation v(m) = m1 + 2m2 + 3m3 , ζ m := n ζnmn ; recall |αj,l | 6 |sjl |
from Lemma 1, k > 1 and
sup |gκ | 6 min{1, const · |j |−r }. (23)
Vj
264 FRANZ MERKL
The case κ > 0 is easier since there is no x-dependence; here we get as well
k
Y Y
ν ` m`
∂j D gκ = |gκk |δ0,m δ0,νκ 0
0
`=1 κ 6=κ
6 constk · |j |−r · δ0,m 6 const0k · |sj |v(m) |j |−r .

Summing the terms up, we get a bound over Vj :
ν m
∂ D f (gκ ) 6 constν,m · |sj |v(m) |j |−r . (24)
j
It results

ν m −1 ∂φl
∂ D φ 6 constν,m,x̂ · |sj |v(m) |j |−r ; (25)
j l
∂z j
we included the case (ν, m) = (0, 0) in the last two bounds too; see the estimates
(21) and (23).
The next lemma examines ϕ = φ0 only; it yields a bound for the left upper
corner M11 in the matrix Mx,P :
LEMMA 4. Assume that |cj ||sj |v(m)+2 |j |−r → 0 as |j | → ∞. Then multiplica-

tion with D m ∂ϕ/ϕ is a compact operator Q V → B̌; it depends continuously in the
operator norm on the divisors P κ ∈ j aj .
Proof. The function ∂ϕ/ϕ is supported in the union of all Vj ; we write it as
P
∂ϕ/ϕ = ±j ∈J 1Vj ∂ϕ/ϕ; here 1Vj means the indicator function of the set Vj . The
P
sum ±j ∈J 0 1Vj D m(ϕ −1 ∂ϕ): V → B̌ over finite subsets J 0 ⊂ J is compactly
supported; hence it is a compact multiplication operator; to see this recall that the
norm k·kV in V controls a first derivative too; see [12] for more details. We estimate
the operator norm of the error term: For f ∈ V ,

X
m ∂ϕ
f · 1Vj D
ϕ
±j ∈J \J 0 B̌
X Z 2 !1/2
|dz | ∂ϕ
6 |f |A j B̌
D m
ϕ −1
|1| ∂z
V A j
±j ∈J \J 0 j

|dzj |B̌ m −1 ∂ϕ
6 sup sup D ϕ kf kV
±j ∈J \J 0 Vj |1|A ∂zj
6 constm,x̂ sup |cj ||sj |v(m)+2 |j |−r kf kV
±j ∈J \J 0
we utilized the bound (25) and the bounds on the Hermitian metrics | · |A and
| · |B̌ . The last supremum goes to 0 as J 0 increases towards J ; this shows that the
multiplication operator with ϕ −1 ∂ϕ can be approximated in the operator norm by
compact operators; hence it is compact.
Q
It remains to examine the dependence on P κ : Let Qκ , R κ ∈ j aj . We join
Q
Qκ and R κ by a path [0, 1] → ( j aj )N , s 7→ (P κ (s))κ , P κ (0) = Qκ , P κ (1) =
R κ , zj (Pjκ (s)) = zj (Qκj )eisqj,κ with |qj,κ | 6 π ; this means that we have para-
metrized the shorter arc on aj joining Qκj and Rjκ . To avoid confusion, we write the
parameters explicitly in the next computation:

f · D m ∂ϕx,Q − D m ∂ϕx,R
ϕx,Q ϕx,R B̌
Z
X 1 ∂ m ∂ϕx,P (s)

= f · qj,κ D ds
∂p ϕ
j,κ 0 j,κ x,P (s) B̌
−r
6 constm,x sup |qj,κ ||cj ||sj |
v(m)+2
|j | kf kV .
j,κ
This time we used the estimate (25) for ν being the κ-th unit vector. Q The last
supremum converges to 0 when R → Q in the product topology on ( j aj )N ,
i.e. qj,κ → 0 for all j ∈ J and κ = 1, . . . , N; this means that D m (∂ϕx,Q /ϕx,Q )
depends continuously on Q. 2
The matrix element M13 is estimated as follows:
−1
LEMMA 5. Assume that v(m) + n − r < −1. Then the norm kD m (ϕx,P ∂ϕl,x,P )kB̌
m −1
is finite. The function D (ϕx,P ∂ϕl,x,P ), viewed as an element of B̌, depends con-
Q
tinuously on the divisors P κ ∈ j aj .
Proof. We estimate ϕl /ϕ over Vj . We work with varying values of l; to avoid
0,l,x
confusion, we have to write the parameter-dependence explicitly: Pj0 = Pj . The
cutoff function χ equals 1 over Vj , hence
X
ϕl −l
= z exp χj,1 · (l · (Log z − αj,log − ηj,log ) + h) (26)
ϕ j
with
0,l,x 0,0,x
zj (Pj ) zj (Pj )
h := Log 1 − − Log 1 − ;
zj zj
one should note that all terms using zj (Pjκ ) with κ > 0 cancel. The ratio (26) is
uniformly bounded:

ϕl
sup sup < ∞ (27)
j Vj ϕ
as a consequence of Lemma 1 and supj supVj |zj (Pjκ )/zj | < 1.

266 FRANZ MERKL
ν
We estimate the derivatives ∂j D m (ϕl /ϕ) over Vj too: One observes that the only
(x, P )-dependent term in the right-hand side of expression (26) is h; its derivatives
are bounded by the estimate (24):
|∂j D mh| 6 constν,m · |sj |v(m) |j |−r .
ν
ν
We apply the chain rule to write the derivatives ∂j D m ϕϕl as a finite sum:
ν ϕl ϕl X Y
k
ν
∂j D m = const(ν ` ),(m` ) ∂j ` D m` h
ϕ ϕ (ν ` ),(m` ) `=1
Pk Pk
with `=1 ν ` = ν and `=1 m` = m; so we get
ϕl X Yk
const(ν ` ),(m` ) (|sj |v(m` ) |j |−r ) 6 constν,m,x̂
ν
|∂j D m |6 (28)
ϕ (ν )(m ) `=1
` `
Q
we have used the estimate (27); note that the product k`=1 may be empty for
m = 0.
We apply the product rule to ϕ −1 ∂ϕl /∂zj and bound the terms by (28) and (25):

ν m −1 ∂ϕl
∂ D ϕ
j ∂zj
X X
ν 00 m00 ϕl ν 0 m0 −1 ∂ϕl
6 constν,ν ,m,m ∂j D ∂ D ϕl
ϕ j ∂zj
0 0
00 0 00
ν +ν =ν m +m =m0
X X 0
6 constν,ν 0 ,m,m0 ,x̂ · |sj |v(m ) |j |−r
ν 0 6ν m0 6m
6 constν,m,x̂ · |sj |v(m) |j |−r . (29)

This implies that
Z Z
m −1 2 2
D ϕ ∂ϕl 6 constm,x̂ · |sj |v(m) |j |−r 2 dzj
B̌ B̌
Vj Vj
6 const0m,x̂ · |sj | 2(v(m)+n) −2r

|j | ;
recall that ∼ |cj−2 sj−4 | d2 zj and |dzj |B̌ ∼ |cj sjn+2 | over Vj . This estimate yields
m −1
D (ϕ ∂ϕl ) 2
Z B̌
XZ
m −1 2
= D ϕ ∂ϕl + D m ϕ −1 ∂ϕl 2
B̌ B̌
supp(1−χ) j Vj
X
6 constm,x̂ + const0m,x̂ |sj |2(v(m)+n) |j |−2r
j
X
6 constm,x̂ + const00m,x̂ |j |2(v(m)+n−r) .
j
The summation index j runs over a two-dimensional lattice; therefore the norm is
finite when v(m) + n − r < −1.
QWe estimate the dependence on the divisors P κ : We use again a path [0, 1] →
( j aj )N , s 7→ (P κ (s)), P κ (0) = Qκ , P κ (1) = R κ , zj (Pjκ (s)) = zj (Qκj )eisqj,κ
with |qj,κ | 6 π as in the proof of the previous lemma. We do not suppress the
parameters x, P here to avoid confusion. Using the estimate (29) this time for ν
being a unit vector, we get similarly to the previous bound:
2
d m −1
D ϕ
ds x,P (s) ∂ϕl,x,P (s)
XB̌
6 constm,x̂ + constm,x̂ |qj,κ |2 |j |2(v(m)+n−r) < ∞
j,κ
−1
when v(m) + n − r < −1. This shows that D m (ϕx,P ∂ϕl,x,P ) depends continuously
Q
on the divisors P ∈ j aj .
κ
2
We define the ansatz functions µj,x,P , j ∈ J by

κ=0,l=0,x
zj · (zj (Oj ) − zj (Pj ))
µj,x,P = χj,2 · ϕ · κ=0,l=0,x
(zj − zj (Oj ))(zj − zj (Pj ))
κ=0,l=0,x
z−j · (z−j (Oj ) − z−j (Pj ))
= −χj,2 · ϕ · κ=0,l=0,x
= −µ−j,x,P ;
(z−j − z−j (Oj ))(z−j − z−j (Pj ))
κ=0,l=0,x
the factor (zj − zj (Pj )) cancels with one factor in ϕ: µj,x,P is a smooth
section in F(O).
The µj,x,P are supported in Uj ; they are holomorphic sections in the line bun-
dle F(O) over the support of χj,3 (on this domain χj,2 = 1). We examine the
κ=0,l=0,x
nondegenerate case Oj 6= Pj first: we calculate:
1
resOj µj,x,P = zj (Oj )ϕ(Oj )resOj .
zj − zj (Oj )
The residue on the left-hand side is calculated in the bundle F, while on the right-
hand side it is calculated in the trivial line bundle; ∂zj (Oj ) := resOj 1/(zj −zj (Oj ))
is viewed as a complex tangential vector at Oj .
We define also a “residue operator” at Oj for sections r ∈ V ;
Z
zj (Oj ) r ∂χj,3
resOj : r 7→ dzj ∧ dzj · ∂zj (Oj );
2π i Vj ∪V−j zj ∂zj
by Stokes formula this new definition coincides with the original residue whenever
r is meromorphic over a neighbourhood of the support of χj,3 with only one pole
of first order allowed at Oj .
268 FRANZ MERKL
We get the element M21 in the matrix M:

Z
1 r ∂χj,3
M21 : V 3 r 7→ dzj ∧ dzj ∈ CJ .
2π i Vj ∪V−j zj ∂zj j ∈J
There is no dependence on x and P κ here; hence this definition makes sense in the
κ=0,l=0,x κ=0,l=0,x
degenerate case Oj = Pj too. In the nondegenerate case Oj 6= Pj
we have

ϕ(Oj ) resOj r
(M21 r)j = .
resOj µj j ∈J
LEMMA 6. M21 : V → `2w (J ) is a bounded linear operator. All its derivatives

with respect to x and P vanish.
Proof. We estimate over Vj :
Z 2
1 r ∂χj,3
dz ∧ dz
2π i j j
Vj zj ∂zj
Z Z
1 ∂χj,3 dzj ∧ dzj 2
6 const1 · |r|A ·
2

Vj Vj |1|A zj ∂zj
Z Z
1 2 4 2 −2 −4 2
6 const2 · |r|A ·
n cj sj |cj sj | d zj
2
Vj Vj sj
Z
6 const · |cj sj |
2 4−2n
|r|2A . (30)
Vj
We note reflection symmetry of the definition of resOj :

zj (Oj ) ∂χj,3
dzj ∧ dzj ⊗ ∂zj (Oj )
zj ∂zj

z−j (Oj ) ∂χj,3
= dz−j ∧ dz−j ⊗ ∂z−j (Oj ).
z−j ∂z−j
We get by using this symmetry a similar estimate over V−j :
Z 2 Z
1 r ∂χ
j,3
dz ∧ dz 6 const · |c 2 4−2n
s | |r|2A
2π i z ∂z
j j j j
V−j j j V−j
and consequently
X
kM21 rk22,w = |cj−2 sj2n−4 | · |(M21 r)j |2
j ∈J
XZ
6 const · |r|2A
j Vj ∪V−j
6 const · krk2V .
This shows that M21 is bounded. 2
We finally estimate the matrix element M12 :

LEMMA 7. Assume that r > 4. Then
X ∂µj
(M12 )x,P : `2w (J ) → B, β 7→ βj
j
ϕ
is a compact operator. Its derivative D m (M12 )x,P exists as aQcompact operator too
when r > v(m) + 4, and it depends continuously on P κ ∈ j aj .
Proof. We use the notation
p ν
Y ∂ νκ
zj (Pjκ ) =: tj eipj,κ , pj ∈ R and ∂j :=
κ
∂pjνκ
once more (case l = 0); the convention (22) is still valid. We let P κ again depend
on a real parameter s with Qbounded values of qj,κ := dpj,κ /ds when we analyze
the dependence on P κ ∈ j aj .
√
∂µj zj · (zj (Oj ) − tj eipj,0 )
= √ · ∂χj,2 .
ϕ (zj − zj (Oj ))(zj − tj eipj,0 )
Taking derivatives of the pieces we get
p p p v(m)
sup |∂j D m (zj (Oj ) − tj eipj,0 )| 6 const0ν,m tj |α m | 6 const00ν,m tj |sj |
ν
Vj
and

ν m X p k m kipj,0
1 α e
sup ∂j D
√ ipj,0 = sup const k,ν,m tj √ ipj,0 k+1
Vj z j − tj e Vj (z j − tj e )
06k 6m
X
const0k,ν,m |j |−rk |sj | 6 const000
v(m) v(m)
6 ν,m |sj |
06k 6m
with m depending on m > 0 and ν > 0. The product rule and supj supVj |zj /(zj −
z(Oj ))| < ∞ imply

ν m ∂µj p
sup ∂j D 6 const(4) tj |s v(m) | sup |∂χj,2 |
ν.m j B̌
Vj ϕ B̌ Vj
6 constν,m |j |−r |cj sj

v(m)+n+2
|
and consequently
2
X
ν ν m ∂µj
βj (iq j,· ) ∂j D
ϕ
j B̌
X Z
−r v(m)+n+2 2
6 kβk22,w · |cj2 sj4−2n | ν,m |j | |cj sj
const(5) |
j Vj
X
|j |−2r |cj2 sj
2v(m)+4
6 const(6)
ν,m kβk2,w ·
2
|.
j
270 FRANZ MERKL
The last sum is finite when v(m) − r + 3 < −1 because of |cj | 6 O(|j |) and
|sj | 6 O(|j |); recall that j runs through a two-dimensional lattice. This shows that
M12 is a bounded operator. Sums over finite subsets of J yield operators of finite
rank; since these partial sums approximate M12 in the operator norm, M12 and its
derivative D m M12 must be compact operators. Using the estimates for ν being the
κ-th unit vector, we see that these operators depend continuously on P κ . 2
FREDHOLM THEORY FOR THE ASYMPTOTIC EXPANSION
We
P summarize the outcome of the previous four lemmas; we use the divisor P :=
κ
n P
j,κ κ j :
PROPOSITION 4. Mx,P : V ⊕ `2w (J ) ⊕ Cn → B̌ ⊕ `2w (J ) is a Fredholm operator

of index 1. Its derivatives D m Mx,P exist for r > v(m) + max{4, n + 1}; they are
bounded operators and depend continuously on the divisor P .
Proof. The Cauchy–Riemann operator ∂: V → B̌ is a Fredholm operator of
index 1 − n (see the index version of the Riemann–Roch theorem for infinite genus
Riemann surfaces in [12]); M11 = ∂ + ϕ −1 ∂ϕ is a perturbation of this operator by
a compact multiplication operator; hence M11 : V → B̌ is a Fredholm operator of
index 1 − n too. M21 : V → `2w (J ) is bounded; consequently

M11 0
: V ⊕ `2w (J ) → B̌ ⊕ `2w (J )
M21 1
is a Fredholm operator of index 1 − n too; to see this one should note that

M11 0 1 0 M11 0
= : V ⊕ `2w (J ) → B̌ ⊕ `2w (J );
0 1 M21 1 M21 1
the first factor is a Fredholm operator with index 1 − n, while the second one
is bounded with a bounded inverse. Including n additional columns increases the
index by n; hence

M11 0 M13
: V ⊕ `2w (J ) ⊕ Cn → B̌ ⊕ `2w (J )
M21 1 0
is a Fredholm operator of index 1. The compact perturbation by M12 does not
change the index; consequently

M11 M12 M13
Mx,P = : V ⊕ `2w (J ) ⊕ Cn → B̌ ⊕ `2w (J )
M21 1 0
is a Fredholm operator of index 1 too. Existence of the derivatives D m and contin-
uous dependence of P of the entries in Mx,P was shown in the lemmas. 2
Let Mx,P (r, β, α)T = 0. We define ψ by the Equation (16). By construction,

the first row in Mx,P guarantees that ψ is a holomorphic section in F(O) since
(ϕx,P ) is a (only smooth) section in F, r is a (only smooth) section in O(O) and µj
is a (only smooth) section in F(O); the pointwise asymptotic behaviour of ψ will

be studied below. Over the support of χj,3 we have rϕx,P = rφj,0 = ψ − βj µj ;
the functions ψ, µj ϕx,P are meromorphic in this domain; hence r is meromorphic
over supp χj,3 too. Only one single pole of r at Oj may be present in this domain
since krkV < ∞. Consequently the value of the extended residue operator resOj r
(see the definition of M21 ) coincides indeed with the residue of r at Oj , as was
already suggested by the name of the operator. This means that the second row
in the equation Mx,P (r, β, α)T = 0 guarantees that there is no pole of ψ at Oj ,
viewed as a section in F. This shows that (ψ) > −P .
We examine the behaviour at infinity of ψ: Here is a rough bound: Over the
domain U ∩ {|z| > 1} outside the handles the functions µj vanish and ϕl coincides
with z−l exz ; therefore
Z Z n−1 2
X

|e−xz ψ|2 = αl z−l + r < ∞ (31)
U ∩{|z|>1} U ∩{|z|>1} l=0
R
since U ∩{|z|>1} |z−l |2 < ∞ and
Z Z
|r|2 6 const |z−2n| · |r|2A
U ∩{|z|>1}
ZU ∩{|z|>1}
6 const |r|2A
U ∩{|z|>1}
6 constkrk2V < ∞.
2
Corollary 3.38 in [12] tells us that this L -condition implies a pointwise condition:
ψ ∈ 0(Ex (P )); this means
lim sup |e−xz ψ| < ∞;
|z|→∞,z∈U
the corollary is applied to the bounded line bundle Ex (P ); we know that Riemann–
Roch theorem for infinite genus Riemann surfaces is valid for this bounded line
bundle. 0(Ex (P )) being one-dimensional implies that the kernel of Mx,P is one-
dimensional too: The kernel is at least one-dimensional since the index of Mx,P
equals 1, but itP cannot be more than one-dimensional:
Pn−1 The latter would imply that
−1 −1
there is a r = j ∈J βj ϕ µj,x,P + l=0 αj ϕ ϕl,x,P ∈ V with (β, α) 6= (0, 0).
If l < n were the smallest index with αl 6= 0 then |r| ∼ |z−l | near infinity over U ;
this contradicts r ∈ V . Consequently α must vanish. But then the βj must vanish
too; else r had a pole at Pj0 (pole of second order in the degenerate case Oj = Pj0 );
this would contradict r ∈ V too.
We show next that Mx,P (r, β, α)T = 0 and α0 = 0 imply (r, β, α)T = 0: In
this case, we can improve the above estimate (31) of the L2 -norm of ψ (defined as
above) by one order:
Z Z n−1 2
X

|ze−xz ψ|2 = αl z1−l + zr < ∞. (32)
U ∩{|z|>1} U ∩{|z|>1}
l=1
272 FRANZ MERKL
Q know that the line bundle Ex (P ) is isomorphic to some Ob (Q) for a Q−1∈
We
j aj ; when ψ 6= 0 such an isomorphism is given by multiplication with ψ ,
since the only global sections in Ob (Q) are constants. But then ψ −1 must fulfill the
bound
lim sup |exz ψ −1 | < ∞;

|z|→∞,z∈U
This lower bound for ψ contradicts the upper L2 -bound (32). We conclude that
only ψ = 0, (r, β, α)T = 0 is possible when α0 = 0. We summarize these
considerations:
PROPOSITION 5. Let π0 denote the linear form π0 (r, β, α)T = α0 . Then

Mx,P
Nx,P := : V ⊕ `2w (J ) ⊕ Cn → B̌ ⊕ `2w (J ) ⊕ C
π0
is a bounded linear operator with a bounded inverse. Set (rn , β, α)T =

−1
Nx,P (0, 0, 1)T . Then the derivatives D m (rn , β, α)T exist for r > v(m) + max{4,
n + 1}; these derivatives are continuous functions of P , and the asymptotic expan-
sion (9) holds.
7. Pointwise Bounds for the Error Term

We derive pointwise bounds for the error term rn over U near infinity. The key to
the proof is the following lemma, which was proved in [12], Proposition 3.37; we
state it only for the special case (line bundle Ex (P ) ⊗ Ob (n∞)) that we need:
LEMMA 8. Let f0 be a holomorphic function over supp χj,0 ∩ {|z| > 1}, fj be
a meromorphic function defined over Uj with at most a simple pole at Pj . Assume
that the bounds
Z
|zn e−xz f0 |2 < ∞ (33)
supp χj,0
and
sup |zne−xz (f0 − fj )| 6 const · |cj−1 sj−2 | (34)

Uj ∩supp χj,0
hold. Then limz→∞,z∈U zn e−xz f0 ∈ C exists.
From this we derive:
PROPOSITION 6. The pointwise limit of the error term limz→∞,z∈U zn D m rn exists

for r > v(m) + max{4, n + 3}.
Proof. We apply an induction over m with varying n: Differentiation of the

asymptotic expansion
!
X
n−1
−l
ψ = e rn +
xz
z αl
l=0
(valid over supp χj,0 ) yields

!
X m X
n+v(k)−1
D m ψ = exz zv(k) · D l rn+v(k) + z−l D l αl (35)
k+l=m
k l l=0
m
with positive constants kl
. We used that
X
n+v(k)−1
D rn = D rn+v(k) +
l l
z−l D l αl
l=n
holds over supp χj,0 for
r > v(m) + max{4, n + 1} > v(l) + max{4, n + v(k) + 1}. (36)
We apply the previous lemma to
X m n+v(k)−1
X
f0 := D ψ − e m xz
zv(k)−l D l αl over supp χj,0 , (37)
k+l=m
k l l=0
X m n+v(k)−1
X
fj := D ψ − m
φj,l−v(k) D l αl over Uj ; (38)
k+l=m
k l l=0
recall the definition (20) of φj,l−v(k) . We calculate
X m n+v(k)−1
X
n −xz
z e (f0 − fj ) = −z n
(zv(k)−l − e−xz φj,l−v(k) )D l αl
k+l=m
k l l=0
X m X n−1
= − zn−l · (1 − zl e−xz φj,l) )D l αl+v(k) .
k+l=m
k l l=−v(k)
Lemma 1 and supVj |zj (Pjκ )/zj | 6 O(|j |−r ) imply
sup |1 − zl e−xz φj,l |

Uj ∩supp χj,0
N
Y zj (Pjκ ) −nκ
= sup 1 − e x̂(−ẑ+η̂+α̂)
· 1− 6 O(|j |−r )
z
Uj ∩supp χj,0 κ=0 j
274 FRANZ MERKL
since l 6 n − 1 6 r + 1; to see this one should examine the two connected

components of Uj ∩ supp χj,0 separately. This implies the hypothesis (34) in the
previous lemma:
X
n−1
sup |zne−xz (f0 − fj )| 6 constm,x,l,(D· α· ) |sjn−l ||j |−r
Uj ∩supp χj,0 l=−v(m)
||j |−r ) 6 O(|j |−3 )

n+v(m)
6 O(|sj
6 O(|cj−1 sj−2 |)
the last steps hold for n + v(m) − r 6 −3. To check the hypothesis (33) one writes
X m
n −xz
z e f0 = zn+v(k) D l rn+v(k)
k+l=m
k l
(see Equations (37), (35)) and uses

Z
|zn+v(k) D l rn+v(k) |2 < ∞;
supp χj,0
the last bound holds when Proposition 5 is applicable, i.e. when the inequality (36)
is valid. Lemma 8 yields that the following limit exists:
X m
n −xz
lim z e f0 = lim zn+v(k) D l rn+v(k) .
z→∞,z∈U z→∞,z∈U
k+l=m
k l
The induction hypothesis tells us that the limits of the summands

lim zn+v(k) D l rn+v(k)
z→∞,z∈U
exists for l 6 m, l 6= m; so limz→∞,z∈U zn D m rn must exist too. 2
8. Derivation of the KP-Equation from the Asymptotic Expansion

We derive relations between the coefficients of the derivatives of the asymptotic
expansion
2 +t z3
ψ := ψx,P = exz+yz (1 + a1 (x, y, t)z−1 + a2 (x, y, t)z−2 +
+ a3 (x, y, t)z−3 + O(z−4 )).
The calculations get simpler when the series is written in the exponent:
2 +t z3 +a(x,y,t )z−1+b(x,y,t )z−2+c(x,y,t )z−3+O(z−4 )
ψ = exz+yz ,
where a = a1 , b = a2 − 12 a12 , and c = 13 a13 − a1 a2 + a3 . The following derivation of
the KP-equation needs pointwise bounds on the error term D m rn for v(m) + n 6 5
and the existence of D m al for v(m) + l 6 5; Propositions 5 and 6 provide these

bounds for r > 8.
We calculate the following derivatives:

ψx = z + ax z−1 + bx z−2 + cx z−3 + O(z−4) ψ,

ψxx = z2 + 2ax + (2bx + axx )z−1 + ax2 + 2cx + bxx z−2 + O(z−3) ψ,

ψxxx = z3 + 3ax z + 3(bx + axx ) + 3ax2 + 3cx + 3bxx + axxx z−1 +

+ O(z−2) ψ,

ψy = z2 + ay z−1 + by z−2 + O(z−3 ) ψ,

ψxy = z3 + ax z + (ay + bx ) + (by + cx + axy )z−1 + O(z−2 ) ψ,

ψt = z3 + at z−1 + O(z−2 ) ψ.
We build linear combinations to eliminate the terms that are growing at infinity:
u · ψ := ψxx − ψy

= 2ax + (axx − ay + 2bx )z−1 + O(z−2 ) ψ; (39)
w · ψ := 2ψt + ψxxx − 3ψxy

= (3axx − 3ay ) + 2at − 3by + 3ax2 − 3axy + 3bxx + axxx z−1 +

+ O(z−2 ) ψ. (40)
This means that u · ψ and w · ψ are bounded multiples of ψ in U near infinity.

u · ψ and w · ψ can only have poles at the divisor P , since the D mψ can have only
poles there. We conclude that u · ψ and w · ψ are global sections in 0(Ex (P ));
hence they are constant multiples of ψ, since ψ generates 0(Ex (P )). This means
that all higher order terms in the asymptotic expansions (39) and (40) of u and w
must vanish:
axx − ay + 2bx = 0, (41)

2at − 3by + 3ax2 − 3axy + 3bxx + axxx = 0; (42)
the expansions (39) and (40) reduce to
u = 2ax , w = 3axx − 3ay ,
and the defining equation (39) for u now turns out to be the heat equation (10).
We take a linear combinations of derivatives of Equations (41) and (42); we
want the dependence on b to cancel. We end up with the KP-I-equation:
0 = 3(axx − ay + 2bx )y − 3(axx − ay + 2bx )xx +

+ 2(2at − 3by + 3ax 2 − 3axy + 3bxx + axxx )x
= 4at x − axxxx − 3ayy + 12ax axx

= 2 ut − 14 uxxx − 32 ayy + 32 ux uxx .
276 FRANZ MERKL
9. Almost Periodicity of the KP Flow

Q Q Q
We show first that the Jacobi group operations ⊕O : j aj × j aj → j aj ,
Q Q Q Q
O : j aj → j aj , and the KP group operation v: R3 × Q j aj → j aj are
continuous maps; we will conclude from the compactness of j aj that the KP
flow is almost periodic:
Q
Let P , Q, O ∈ j aj and x ∈ R3 . Recall that the divisors P ⊕O Q, O P , re-
spectively v(x, P ) are the zero divisors of ψ0,P +Q−O , ψ0,2O−P , respectively ψx,P ;
these functions are viewed here as sections in Ob (P + Q − O), Ob (2O − P ),
respectively Ex (P ). We conclude from Proposition 5 that the sections ψ0,P +Q−O ,
ψ0,2O−P , ψx,P depend continuously on the parameters (e.g., with respect to the
norm k · kV taken for the error term rn ); hence their zeros depend continuously
on the parameters too. (To see this, one may use a Cauchy-type integral formula,
integrating over a neighbourhoodQ of aj , to express the coordinates of the zeros.)
Q
Choose any metric d on Q j aj that generates the product topology on j aj .
By Tychonov’s theorem, j aj is a compact topological space, so ⊕O , O are
uniformly continuous. Q
Let ε > 0. For P ∈ j aj let uP denote the solution of the KP-equation that
was constructed Q in the last section. We know from Proposition 5 and from the
m
compactness of Q j aj that the derivatives D umP (0) depend muniformly continu-
ous on P ∈ j aj for v(m) < r − 5, say |D uX (0) − D uY (0)| < ε when
d(X, Y ) < ε1 . By uniform continuity of ⊕O , choose δ > Q 0 so that d(X, Y ) < δ
implies d(Z ⊕O X, Z ⊕O Y ) < ε1 for all QX, Y, Z ∈ j aj . We can cover the
orbit {v((0, 0, t), O) | t ∈ R} of O in j aj by finitely Q many sets Wn , n =
1, . . . , N of diameter less than δ; here the compactness of j aj is used again.
We may assume that every Wn has a nonempty intersection with the orbit. Choose
L > 0 so large that every Wn contains a v((0, 0, tn ), O) with −L/2 6 tn 6
L/2. Let [a − L/2, a + L/2] be an interval of length L. Then there is at least
one Wn containing v((0, 0, a), O). But v((0, 0, tn ), O) ∈ Wn too, consequently
Q 0, a), O)) < δ. We 3set τ := a − tn ∈ [a − L/2, a + L/2].
d(v((0, 0, tn ), O), v((0,
Then for every P ∈ j aj and (x, y, t) ∈ R :
d(v((x, y, t), P ), v((x, y, t + τ ), P ))

= d(v((x, y, t − tn ), P ) ⊕O v((0, 0, tn ), O), v((x, y, t − tn ), P ) ⊕
⊕O v((0, 0, a), O)) 6 ε1
and consequently
|D m uP (x, y, t) − D m uP (x, y, t + τ )|
= |D m uv((x,y,t ),P )(0) − D m uv((x,y,t +τ ),P )(0)| < ε.
10. Bloch Condition, Constraint, and Initial Condition

We use the notation of Section 2 again.
The projection maps 4n : H(q) 3 (ξ1 , ξ2 ) 7→ ξn ∈ C∗ , n = 1, 2, are nowhere
vanishing holomorphic maps on the heat curves with the asymptotic behaviour
|z(ξ )|→∞
4n (ξ )e−(γn1 z(ξ )+γn2 z(ξ )
2)
−→ 1;
here γn = (γn1 , γn2 ) is a basis vector in the lattice 0; see the asymptotic description
of heat curves (8) resulting from [2], Part III, §16, Theorem 16.1. This means that
4n ∈ 0(Eγn ); we have set γn := (γn1 , γn2 , 0). Multiplication with 4n defines
therefore an isomorphism Ex (P ) → Ex+γn (P ) of bounded line bundles. This
isomorphism also acts on the one-dimensional space of global sections:
4n · ψx,P = ψx+γn ,P ; (43)
note that both sides are normalized by their behaviour at infinity. The last equation
just states the Bloch condition (6).
Let
4n = eγn1 z+γn2 z (1 + aγn z−1 + O(z−2 )),

2
2 +t z3
ψx,P = exz+yz (1 + a(x)z−1 + O(z−2 ))
denote the asymptotic expansions in U near infinity; then the Bloch condition
(43) yields aγn + a(x) = a(x + γn ). Differentiation with respect to x implies
0-periodicity of the solution of the KP-I equation: u(x) = u(x + γn ) and that the
mean value in x-direction of u(x) equals a constant:
Z γ11
u(x + s, y, t) ds = 2(a(x + γ11, y, t) − a(x, y, t)) = 2aγ1 ;
0
this integral must vanish when it vanishes for the initial conditions u(x, y, t = 0).
Let ψξ (x, y) 6= 0 solve the heat equation (5) with the Bloch condition (6). Then
we can reconstruct the potential q by
q(x, y) = ψξ (x, y)−1 (∂x2 − ∂y )ψξ (x, y). (44)

Q
We can rephrase the content of Proposition 2 as Fx,y ∈ 0(E(x,y,0)(P )); P ∈ j aj
denotes the zero divisor of H(q) 3 ξ 7→ (Fξ (q) 3 ψξ 7→ ψξ (0, 0)). The space
0(E(x,y,0)(P )) is generated by the section ψ(x,y,0),P ; the latter has the same as-
ymptotic behaviour as Fx,y ; consequently Fx,y = ψ(x,y,0),P . This implies the initial
condition u(x, y, 0) = q(x, y) since
−1
u(x, y, 0) = ψ(x,y,0),P (∂x2 − ∂y )ψ(x,y,0),P ;
see Equation (39).

278 FRANZ MERKL
Acknowledgements
The author wants to thank Horst Knörrer and Henry McKean for their continuous
support, their advice and many discussions. The author also would like to thank
Giovanni Felder, Joel Feldman, Fritz Gesztesy, Igor Krichever, Alain-Sol Sznitman
and Eugene Trubowitz for interesting discussions. This work was supported by the
Swiss National Science Foundation, Bern and by the Deutsche Forschungsgemein-
schaft, Bonn.
References
1. Bourgain, J.: On the Cauchy Problem for the Kadomcev–Petviashvili equation, Geom. Funct.
Anal. 3 (1993), 315–341.
2. Feldman, J., Knörrer, H., and Trubowitz, E.: Riemann surfaces of infinite genus I–IV, ETH
Preprint, Zürich, 1994.
3. Feldman, J., Knörrer, H., and Trubowitz, E.: Infinite genus Riemann surfaces, In: J. Carell and
R. Murty (eds), Canadian Mathematical Society 1945–1995, Volume 3, 1996, pp. 91–111.
4. Gesztesy, F., Holden, H., Saab, E., and Simon, B.: Explicit construction of solutions of the
modified Kadomcev–Petviashvili equation, J. Funct. Anal. 98(1) (1991), 211–228.
5. Gesztesy, F. and Schweiger, W.: Rational KP and mKP-solutions in Wronskian form, Rep.
Math. Phys. 30(2) (1991), 205–222.
6. Gesztesy, F., and Unterkofler, K.: On the (modified) Kadomcev–Petviashvili hierarchy, Differ-
ential Integral Equations 8(4) (1995), 797–812.
7. Krichever, I. M.: Spectral theory of two-dimensional periodic operators and its applications,
Russian Math. Surveys 44(2) (1989), 145–225.
8. McKean, H. P.: Integrable systems and algebraic curves, In: Global Analysis, Lecture Notes in
Math. 755, Springer, New York, 1978, pp. 83–200.
9. McKean, H. P. and Trubowitz, E.: Hill’s surfaces and their theta functions, Bull. Amer. Math.
Soc. 84(6) (1978), 1042–1085.
10. McKean, H. P. and Trubowitz, E.: Hill’s operator and hyperelliptic function theory in the
presence of infinitely many branch points, Comm. Pure Appl. Math. 29 (1976), 143–226.
11. McKean, H. P. and van Moerbeke, P.: The spectrum of Hill’s equation, Invent. Math. 30 (1975),
217–274.
12. Merkl, F.: A Riemann–Roch theorem for infinite genus Riemann surfaces, submitted to Invent.
Math. (1999).
13. Merkl, F.: A Riemann–Roch theorem for infinite genus riemann surfaces with applications to
inverse spectral theory, Dissertation ETH Zürich No. 12469, 1997.
14. Mumford, D.: An algebro-geometric construction of commuting operators and of solutions to
the Toda lattice equation, Korteweg–de Vries equation and related non-linear equations, In:
Proceedings of the International Symposium on Algebraic Geometry, Kyoto, 1977, pp. 115–
153.
279
Lifshitz Asymptotics via Linear Coupling

of Disorder
PETER STOLLMANN
Department of Mathematics, Johann Wolfgang Goethe-Universität, Frankfurt, Germany
(Received: 22 December 1998)

Abstract. We present a simple method for proving Lifshitz asymptotics for random Schrödinger
operators and apply it to the Anderson and Poisson model.
Mathematics Subject Classifications (1991): 82B44, 47B80, 60H25.
Key words: Lifshitz tails, integrated density of states.
1. Introduction
In this note, we present a very simple approach to proving Lifshitz asymptotics
for random operators and apply it to Schrödinger operators with Anderson and
Poisson potentials. Let us first briefly recall what Lifshitz asymptotics is about:
consider H0 = −1 on Rd and a random potential term Vω . The operator H0 + Vω
is to be thought of as the effective Hamiltonian of fixed a realization of a random
solid. If we assume that the Vω are bounded below uniformly in ω, Vω > 0 say,
the restriction (H0 + Vω )3 of H0 + Vω to an open cube 3 with Neumann boundary
conditions has compact resolvent. Therefore, the spectral counting function

n(E, (H0 + Vω )3 ) := tr χ[0,E] (H0 + Vω )3
which gives the number of eigenvalues below E, counted with multiplicity, is fi-
nite. This function bears important information about the random potential under
consideration. In fact, its limit as 3 exhausts the whole space has an asymptotic
behaviour characteristic of disorder. To see that, let us first recall that by the cel-
ebrated Weyl asymptotic formula, in absence of disorder, i.e. for Vω = 0, we
have
n(E, (H0 )3 ) = Cd E d/2(|3| + o(|3|)) E > 0
(where we use |3| for the volume of the cube) which means that
1
lim n(E, (H0 )3 ) =: N0 (E) = Cd E d/2 .
3%Rd |3|
280 P. STOLLMANN
A submultiplicative ergodic theorem implies that the respective limit

1
N(E) := lim n(E, (H0 + Vω )3 )
3%Rd |3|
exists for a.e. ω under some mild and very natural ergodicity assumption on Vω
(see [4]). Moreover, this limit is independent of the choice of ω outside some set
of measure zero and equals
1
N(E) = inf E{n(E, (H0 + Vω )3 )}.
3 |3|
Now the right-hand side above is readily interpreted as the expected number of
energy levels per unit volume below E. Clearly, this quantity is of importance
both mathematically and from the physicists point of view. In his landmark work,
Lifshitz predicted an asymptotic behaviour of N(E) which differs drastically from
the dimension-dependent power law decay of N0 (E). Namely, he claimed that
for nontrivial Vω which obeys some spatial independence (this will be explained
below),
N(E) ∼ exp(−γ E −d/2) as E & 0.
(Here we assume that 0 is the inf of the spectrum of H0 + Vω a.e. for notational
convenience.) His reasoning is as follows:
first of all n(E, (H0 + Vω )3 ) 6 n(E, (H0 )3 ) as the nonnegative potential term
shifts the eigenvalues to the right. Therefore, with E1 (. . .) denoting the bottom
eigenvalue of the operator in question, we have
Z
1
N(E) 6 n(E, (H0 + Vω )3 )χ{E1 ((H0 +Vω )3 )6E} dP(ω)
|3|
6 CE d/2 P{ω : E1 ((H0 + Vω )3 ) 6 E}.
Now we want to estimate the probability of having small eigenvalues. If φ is a

normalized eigenfunction of (H0 +Vω )3 with eigenvalue E ∼ 0 it must be localized
to a region where Vω = 0, as E = (−1φ|φ) + (Vω φ|φ). As the kinetic energy of
a function localized to a set of diameter R is at least of order R −2 , there must be a
ball of radius E −1/2 on which Vω vanishes essentially. The spatial independence
referred to above means that we assume that the restrictions of Vω to disjoint
subsets are independent of each other. In that case, the probability that Vω vanishes
on a ball of radius R goes to zero exponentially in the volume R d of the ball as R
goes to infinity. Inserting the length R = E −1/2 found above, we get that
P{ω : E1 ((H0 + Vω )3 ) 6 E} 6 const exp(−γ E −d/2).
Of course, this is not a mathematically rigorous proof. The point which certainly
has to be made precise is the existence of a large enough region where Vω = 0.
LIFSHITZ ASYMPTOTICS VIA LINEAR COUPLING OF DISORDER 281
For due to tunneling effects, φ might still live on parts of space where Vω 0. Of
course, Vω may not increase the potential energy too much.
Our way around that difficulty goes as follows: Let H (ω) = H0 + Vω be a
random Schrödinger operator with Vω > 0. By what we said above, Lifshitz be-
haviour for the integrated density of states can be deduced from an estimate of the
following form, where H3 (ω) denotes H0 +Vω in L2 (3), with Neumann boundary
condition (b.c.), 3 = 3l (0) an open cube with sidelength l in Rd and E1 (·) the first
eigenvalue.
P{E1 (H3 (ω)) 6 E1 (H0 ) + bl −2 } 6 4 exp(−l d γ ). (1.1)
This latter inequality states in precise terms that it is very unlikely to find really
small eigenvalues of H3 (ω). In order to prove such an inequality, one has to over-
come the following main problem: for a simple-minded lower bound on the first
eigenvalue one would need a uniform lower bound on the perturbation Vω . Such
uniform lower bounds only hold with small probability. On the other hand, what
one knows by standard probabilistic tools are lower bounds for the mean of Vω
for typical ω. So what we need is a relation between the mean of Vω and the first
eigenvalue. In our approach we choose a derivative related with E1 (ω) as such a
link. As you will see, that provides a conceptually simple proof of inequality (1.1).
More precisely, let us consider
H3 (ω, t) = H0 + tVω on L2 (3) with Neumann b.c.
Then the first eigenvalue E1 (ω, t) of this operator behaves like
E1 (H3 (ω)) > E1 (ω, t) ≈ E1 (H0 ) + tE10 (ω, 0) for small t, (1.2)
with
E10 (ω, 0) = (Vω φ0 |φ0 ),
where φ0 is the normalized ground state of H0 .
Now, let us take a closer look at (1.2). We have to find out just how large we
may take t. Analytic perturbation theory suggests t ≈ l −2 , as this is the distance of
E1 (H0 ) to E2 (H0 ) for typical Schrödinger operators. With this choice and b small
enough, from E1 (ω) 6 E1 (H0 ) + bl −2 it follows that E10 (ω, 0) has to be small.
But, in the Anderson case,

1 X
E10 (ω, 0) = (Vω φ0 |φ0 ) = ω(i) const. (1.3)
|3| i∈3
is essentially the mean of a sum of |3| = l d i.i.d. variables. The probability that
this mean differs from the expectation by some fixed constant goes to zero expo-
nentially in the number of independent copies, i.e. exp(−γ |3|) which is exactly
the decay we need. For Poisson potentials, we provide a rather elementary large
deviation estimate, reducing it to something like (1.3). The above considerations
282 P. STOLLMANN
constitute already the main idea of our method which we call ‘linear coupling of
disorder’ for obvious reasons. The rest of the paper is devoted to carrying out the
details needed to turn the above heuristics into a rigorous proof.
In principle, this requires three steps: firstly, the standard procedure to deduce
Lifshitz tails from (1.1) above. To prove (1.1) we need, of course, large deviation
results for (Vω φ0 |φ0 ). As a last ingredient, a remainder estimate for the first-order
approximation to E1 (ω, t) is needed, which, again, is quite standard.
As all three steps are rather straightforward, the expert reader could stop at
this point. (But please, read on.) To appreciate the simplicity of our approach, the
reader should compare it with the proofs available so far; see [1, 10], where detailed
references to the literature can be found. Usually, there is some tricky part when
it comes down to showing the main point: small E1 (ω) come from large deviation
from the typical ω.
For the Anderson model, Temple’s or Thirring’s inequality is used at that point.
The Poisson model was treated using the celebrated work of Donsker and Varad-
han [2] on the asymptotics of the Wiener sausage. A beautiful introduction to this
circle of ideas can be found in [8].
In our approach we single out a very convenient link, namely E10 (ω, 0).
This enabled us to apply our method to a random quantum waveguide model [5],
quite reminiscent of the Anderson model. For this model, however, determining the
derivative E10 (ω, 0) is harder (and more interesting), and the methods using Tem-
ple’s or Thirring’s inequality fail. Despite of all the advertisement for our method,
we should stress that, so far, we haven’t achieved ‘the right constant’ in the Poisson
case.
2. Lifshitz Tails for Anderson and Poisson Models

Let us first fix the notation and the basic assumptions. Throughout the following,
3 = 3l (x) denotes an open cube of sidelength l centered at x. Moreover, p is an
exponent such that p > d/2 for d > 4 and p = 2 for d 6 3, and f ∈ Lp (Rd ),
f > 0 and supp f ⊂ 31 (0). We will consider the following random potentials:
(A) THE ANDERSON MODEL. LetR S > 0 and µ be a probability measure on

[0, S] with 0 ∈ supp µ and M = x dµ > 0, i.e. µ is not just δ. Denote its
R d d
variance
P by v = x 2 dµ. Let = [0, S]Z , P = µZ and define VωA (x) =
i∈Zd ω(i)f (x − i).
d
(P) THE POISSON MODEL. Let denote the point P measures on R and P the
Poisson measure on , which is concentrated on { i δXi ; for some discrete se-
quence (Xi )}. Define
X
VωP (x) = f ∗ ω(x) = f (x − Xi (ω)).
i
These two random measures correspond to quite different types of disorder:

while VωA is used to model solids with defects, with some periodicity still present
in the random potential, VωP describes an amorphous medium in which the nuclei
(at Xi (ω)) are distributed erratically in space.
As many of the following considerations apply to both the Anderson and the
Poisson model, we will often write Vω to denote either of them, and use a super-
scripts A, P to distinguish between them. With this convention, denote H (ω) =
−1 + Vω in L2 (Rd ) and by H3 (ω) the restriction of this operator to L2 (3) with
Neumann boundary conditions.
The integrated density of states for H (ω) is given by
1
N(t) = inf E tr χ[0,t ] (H3 (ω))
3 |3|
1
= lim tr χ[0,t ] (H3 (ω)) P-a.s.
3%Rd |3|
We refer to [1, 4, 10] for a discussion of this very important quantity. Note that the
trace appearing above simply counts the number of eigenvalues below t, so that
N(t) is interpreted as the number of energy levels per unit volume of H (ω). The
fact that
1
E tr χ[0,t ] (H3 (ω))
|3|
decreases as 3 increases is due to our choice of the boundary condition. Since
we are working with Neumann boundary conditions the spectral counting func-
tion is subadditive on disjoint open sets. It is also possible to work with Dirichlet
boundary conditions instead, in which case the spectral counting function is super-
additive. For reasonably well defined Vω the limits are in fact the same. See [4] for
a thorough discussion of this point.
The estimate given in the next theorem is usually referred to as Lifshitz tail
behaviour and is one of the central topics of disordered systems ever since Lifshitz’
seminal contribution [9]:
THEOREM 2.1. The integrated density of states N(t) satisfies
log N(t)
lim sup −d/2 6 −γ (2.1)
t &0 t
for some γ > 0. For the Anderson model, γ = γA depends upon f, M, S and for
the Poisson model, γ = γP depends upon f .
The inequality (2.1) will easily follow from the next result, as we will show at
the end of this section. In the supplement given there one can see the dependence
quite clearly. Note that we write E1 (·) for the first eigenvalue of the operator in
question.
PROPOSITION 2.2. (A) There exist universal constants c, K > 0 such that with
cA = c · S · kf kp /kf k1 , for every
284 P. STOLLMANN

π 2 M2
b 6 min , ,
4 cA2
we have
P{E1 (H3A (ω)) 6 b · l −2 }
√ √
d M − cA b S · (M − cA b)
6 K exp −l log 1 + . (2.2)
K ·S v
(P) There exists a universal constant c0 such that, for M = (e − 1)/e = v,
kf kp
cP = c0 · ,
kf k1
and every

π 2 M2
b 6 min , ,
4 cP2
we have
P{E1 (H3P (ω)) 6 b · l −2 }
ld
√ ld
√
M − cP (l−2) d b M − cP (l−2) d b
6 K exp −(l − 2) d
log 1 + .(2.3)
K v
Let us first single out an important step in the proof of Proposition 2.2. To this
p
end, fix V ∈ Lloc,unif(3), let H (t) = −1 + t · V in L2 (3) with Neumann b.c. and
denote its first eigenvalue by E1 (t). Note that E1 (0) = 0.
LEMMA 2.3. There exists a universal constant C such that for τ = C·kV k−1
p,loc,unif
and 0 6 t 6 τ l −2 we have
π2 2 2
|E1 (t) − tE10 (0)| 6 ·l ·t .
4τ 2
Proof. To estimate the remainder term in the Taylor expansion we want to use
[3], formula II(3.6). The isolation distance ϑ defined as the distance of E1 (0) to
the rest of the spectrum of H (0) is given by ϑ = π 2 / l 2 .
As 0 we choose a circle around E0 with radius ϑ/2. We need an estimate for
the r0 appearing in [3], II(3.3), which means that we have to consider
r(ζ ) = kV (H (0) − ζ )−1 k−1 for ζ ∈ 0.
As (H (0) − ζ )−1 maps L2 to the Sobolev space W 2,2 with norm controlled by
dist(ζ, σ (H (0))) > ϑ/2, we have by Sobolev’s inequality that
2
kV (H (0) − ζ )−1 k 6 c0 · · kV kp,loc,unif,
ϑ
so that
ϑ
r0 = min r(ζ ) > C 00 ,
ζ ∈0 kV kp,loc,unif
and an appeal to [3], estimate II(3.6) finishes the proof. 2
We are now ready to present the proof of Proposition 2.2.
Proof of Proposition 2.2. Denote H (ω, t) = −1+t·Vω in L2 (3) with Neumann

b.c. and denote by E1 (ω, t) its first eigenvalue. From the remainder estimate in
Lemma 2.3 we have
π2 2 2
|E1 (ω, t) − t · E10 (ω, 0)| 6 ·l ·t (0 6 t 6 τ l −2 ),
4τ 2
where
1 1
τ =C· >C·
kVω kp,loc,unif S · kf kp
is bounded away from 0, independently of ω. Assume that E1 (ω) 6 b · l −2 for
b 6 π 2 /4. Then the above inequality yields
π2 2 2
t · E10 (ω, 0) 6 · l · t + b · l −2 for all 0 6 t 6 τ l −2 .
4τ 2
Inserting t = sτ l −2 we get
π 2s b
E10 (ω, 0) 6 + for all 0 6 s 6 1.
4τ τs
√
Optimizing w.r.t. s we get s = 2
π
b and
π√
E10 (ω, 0) 6 b,
τ
which implies
π√
(Vω φ0 |φ0 ) 6 b,
τ
where
1
φ0 = χ3 .
|3|1/2
We now specialize to the case (A):
Then

1 X
(Vω φ0 |φ0 ) = kf k1 ω(i) .
|3| i∈3
286 P. STOLLMANN
Define
π · S · kf kp
cA = ,
C · kf k1
so that
π
6 cA
kf k1 · τ
Now, if 0 6 b 6 M 2 /cA2 it follows that

−2
1 X √
P E1 (ω, 1) 6 b · l 6 P ω(i) 6 cA b
|3| i∈3

1 X √
6 P ω(i) − M > M − cA b .
|3| i∈3
By [11], Thm. 1.4, this latter probability can be estimated by

√ √
M − cA b S · (M − cA b)
K exp − l d log 1 + ,
K ·S v
the assertion.
To treat case (P), we want to use a similar calculation and subdivide 3 into the
unit cubes 31 (m), where m runs through 3l (0) ∩ Zd .
We introduce the random variables

1, if there is an Xi (ω) ∈ 31 (m),
Ym (ω) :=
0, else,
for m ∈ 3l−2 (0). By the properties of the Poisson process, these r.v. are i.i.d. with
expectation and variance equal to M = (e − 1)/e = v. We define an auxiliary
random potential by
X
Wω (x) := Ym (ω) · f (x − Xi(m) (ω)),
m∈3l−2 (0)
where Xi(m) (ω) is one of the Poisson points in 31 (m), if Ym (ω) = 1, and zero else.
Clearly,
Wω (x) 6 Vω (x) for all ω ∈ , x ∈ Rd ,
and, hence,
P{E1 (H3 (ω)) 6 b · l −2 } 6 P{E1 (−1 + Wω ) 6 b · l −2 }.
Now the latter probability can be estimated by the same calculation as in the case
(A) above, since
X
(l − 2)d 1
(Wω φ0 |φ0 ) > kf k1 Ym ,
ld (l − 2)d m
and
kWω kp,loc,unif 6 3d · kf kp ,
where for the last inequality we counted the neighbouring boxes and thus the max-
imal number of nontrivial terms in the sum which defines Wω . Thus, we are again
left with applying a large deviation result for sums of i.i.d. variables. We get, with
π kf kp
cP = · ,
C kf k1
that

P E1 (ω, 1) 6 b · l −2
X
1 ld √
6P Ym 6 cP b
(l − 2)d m∈3 (l − 2)d
l−2

1 X ld √
6P
Ym − M > M − cP b
|3| m∈3 (l − 2)d
l−2
ld
√ ld
√
M − cP (l−2) d b M − cP (l−2) d b
6 K exp − (l − 2) d
log 1 + ,
K v
by [11], Thm. 1.4. 2
It remains to prove the theorem. In order to give more precise information on

the exponent, let us introduce some notation. Denote
√ √
d/2 M − c∗ b S · (M − c∗ b)
γ∗ (b) = b log 1 +
K ·S v
for ∗ = A, P , where M, v, S are defined in (A) for the Anderson case and M =
v = (e − 1)/e, S = 1 in the Poisson case.
SUPPLEMENT TO THEOREM 2.1. Inequality (2.1) holds for
2
π M2
γ∗ = max γ∗ (b); 0 6 b 6 min , .
4 c∗
Proof. We first deduce inequality (1.1) with a b-dependent exponent γ (b), where
b is as in Proposition 2.2. The result above will then follow by optimizing with
respect to b.
First note that

1
N(t) = inf E tr χ[0,t ] (H3 (ω))
3 |3|
1
6 inf P {E1 (H3 (ω)) 6 t} · c · |3|
3 |3|
288 P. STOLLMANN
by Weyl’s law, referred to in the introduction; choosing t = b · l −2 with

2
π M2
0 6 b 6 min , , and 3 = 3l (0),
4 cA2
we get the assertion. 2
3. Concluding Remarks
Of course, one could shorten the above proof if one isn’t interested in the exponent.
There are different quite easy perturbation theoretic proofs for Lifshitz tail
asymptotics which use Temple’s or Thirring’s inequality. See [4] for a detailed
explanation and references. The method presented here has the advantage that the
link between spectral and probability theory provided by the derivative allows for
a conceptually more transparent proof, at least in our opinion. Moreover, the deriv-
ative is in many cases easy to calculate or at least easy to guess, which provides a
road map for the rigorous proof. An example is the application of the above method
in [5], where we didn’t see how to use the methods previously available.
So far we haven’t been able to strengthen our arguments so as to obtain the
d/2
correct value of the exponent which is known to be C · γd , where C is a known
constant and γd is the lowest eigenvalue of the Dirichlet Laplacian on a ball of unit
volume in Rd ; see [10] for an extensive discussion. This correct value is related
with isoperimetric inequalities and is obtained by using the celebrated results of
Donsker and Varadhan; see [2, 8, 4, 10].
To date there are more detailed results available for the bottom (i.e. principal)
eigenvalue of a Schrödinger operator with Poissonian obstacles; we refer the reader
to [12] and the literature cited there.
A recent thorough investigation of the attractive Poissonian case (i.e. the case
where f is nonpositive) is given in [7].
Let us further mention recent deep work of Klopp, [6], which deals with what
is called internal Lifshitz tails. Consider H0 = −1 + V0, where V0 is a Zd -periodic
potential, and H (ω) = H0 +VωA . The spectrum of H0 consists of a number of closed
intervals, called bands, separated by open intervals called gaps. The same is true for
H (ω), where the bands are usually shifted and somewhat enlarged, depending on
sign and size of VωA . Lifshitz predicted that the behaviour of the integrated density
of states N0 for H0 near band edges E0 should have the same power law decay as
in the case V0 = 0 at energy 0, i.e.
N0 (E0 + ε) ∼ N0 (E0 ) + ε d/2 as ε & 0
if E0 is the left endpoint of one of the bands. Moreover, for the randomized op-
erator H (ω) he claimed that the integrated density of states N should exhibit
the exponential decay discussed above for the inf of the spectrum. Interestingly
enough, both claims are still not proved nor disproved in general. Klopps work
establishes an equivalence between them, saying that at band edges at which N0

behaves as predicted so does N. Presumably, this can be proven by our methods
above (work in progress, jointly with G. Stolz). However, Klopp’s article gives
more information: expanding the projection of the periodic Hamiltonian onto a
band into Wannier functions he establishes a certain equivalence with a discrete
Schrödinger type operator. This equivalence is used to reduce the proof of Lifshitz
asymptotics near band edges to the proof of Lifshitz asymptotics of an associated
discrete operator at the bottom of the spectrum.
Acknowledgement
Heartfelt thanks go to F. Kleespies and R. Lang for most useful discussions.
References
1. Carmona, R. and Lacroix, J.: Spectral Theory of Random Schrödinger Operators, Birkhäuser,
Boston, 1990.
2. Donsker, M. D. and Varadhan, S. R. S.: Asymptotics for the Wiener sausage, Comm. Pure Appl.
Math. 28 (1976), 525–565.
3. Kato, T.: Perturbation Theory for Linear Operators, Springer-Verlag, New York, 1980.
4. Kirsch, W.: Random Schrödinger operators: A course, In: H. Holden and A. Jensen (eds),
Schrödinger Operators (Sonderborg DK, 1988), Lecture Notes in Phys. 345, Springer, Berlin,
1989.
5. Kleespies, F. and Stollmann, P.: Lifshitz asymptotics and localization for random quantum
waveguides, Rev. Math. Phys., to appear.
6. Klopp, F.: Internal Lifshitz tails for random perturbations of periodic Schrödinger operators,
Duke Math. J. 98(2) (1999), 335–396.
7. Klopp, F. and Pastur, L.: Lifshitz tails for random Schrödinger operators with negative singular
Poisson potential, Preprint, 1998.
8. Lang, R.: Spectral Theory of Random Schrödinger Operators. A Genetic Introduction, Lecture
Notes in Math. 1498, Springer, New York, 1991.
9. Lifshitz, I. M.: Energy spectrum structure and quantum states of disordered quantum systems
(in Russian), Uspekhi Fiz. Nauk 83 (1964), 617–663.
10. Pastur, L. and Figotin, A.: Spectra of Random and Almost-Periodic Operators, Springer-Verlag,
Berlin, 1992.
11. Talagrand, M.: New concentration inequalities in product spaces, Invent. Math. 126(3) (1996),
505–563.
12. Sznitman, A.-S.: Fluctuations of principal eigenvalues and random scales, Comm. Math. Phys.
189 (1997), 337–363.
291
Sharp Spectral Asymptotics and Weyl Formula for

Elliptic Operators with Non-smooth Coefficients
LECH ZIELINSKI
Institut de Mathématiques de Paris-Jussieu UMR9994, Université Paris 7 (D. Diderot), 2 Place
Jussieu, 75252 Paris Cedex 05, Case Postale 7012
(Received: 6 April 1999; in final form: 10 September 1999)

Abstract. The aim of this paper is to give the Weyl formula for eigenvalues of self-adjoint el-
liptic operators, assuming that first-order derivatives of the coefficients are Lipschitz continuous.
The approach is based on the asymptotic formula of Hörmander’s type for the spectral function
of pseudodifferential operators having Lipschitz continuous Hamiltonian flow and obtained via a
regularization procedure of nonsmooth coefficients.
Mathematics Subject Classification (1991): 35P20.
Key words: spectral asymptotics, Weyl formula, self-adjoint elliptic operators with nonsmooth co-
efficients, sharp remainder estimates.
1. Introduction
This paper is motivated by the well-known result of L. Hörmander [6] concerning
the asymptotic formula for the spectral function (i.e. the kernel of the spectral
projector) of a self-adjoint elliptic differential operator A with smooth coefficients
on a smooth manifold M. If 2m is the order of A and d is the dimension of M, then
d
e(A, y, y, λ) = ω(y)λ 2m (1 + O(λ−µ )) (1.1)
holds with µ = 1/2m. If M is compact, then (1.1) implies the Weyl formula
N(A, λ) = ωλ 2m (1 + O(λ−µ )),

d
(1.2)
where N(A, λ) is the counting function of A (i.e. the number of eigenvalues less
than λ, counted with multiplicities).
It is natural to ask if (1.2) still holds for elliptic operators with irregular coeffi-
cients. In [29] we proved that (1.2) holds with µ < r/2m if the coefficients of A are
Hölder continuous of exponent r ∈ ]0; 1]. In particular, for Lipschitz continuous
coefficients, the exponent µ can take an arbitrary value strictly smaller than 1/2m
and we called these results ‘intermediate remainder estimates’.
The aim of this paper is to obtain ‘sharp remainder estimates’, i.e. with the
optimal exponent µ = 1/2m (the value µ = 1/2m cannot be improved in general
292 LECH ZIELINSKI
without additional hypotheses on the Hamiltonian flow). Our basic result in this
direction is
THEOREM 1.1. Let M be a compact (boundaryless) smooth manifold of dimen-

sion d with a density dx and A a self-adjoint operator in L2 (M, dx). Assume that A
is a differential elliptic operator of order 2m of the general form described in The-
orem 1.2 of [29]. If first-order derivatives of coefficients of A (in local coordinates)
are Lipschitz continuous, then the Weyl formula (1.2) holds with µ = 1/2m.
A similar problem of spectral asymptotics can be considered for a manifold with

a boundary. Boundary-value problems are considered in [30] and our basic result
is the estimate
d −1
N(A, λ) = ωλ 2m (1 + O(λ 2m log λ)) (1.3)
for elliptic operators of order 2m > d if first-order derivatives of coefficients are
Lipschitz continuous. We note that in the case of smooth coefficients, the estimate
(1.3) was obtained by Z. Brüning [4] and we refer to [9, 10, 17] concerning the
boundary-value problems in the case of smooth coefficients and to [16, 18 – 21, 23,
25] in the case of irregular coefficients.
Another situation is described in [31], where we consider a class of transitive,
ergodic, elliptic differential operators in Rd (e.g. operators with almost periodic
coefficients). If first-order derivatives of coefficients are uniformly Lipschitz con-
tinuous and 2m > d, then we can define the integrated density of states N̄(λ)
satisfying
d −1
N̄(λ) = ωλ 2m (1 + O(λ 2m )).
Concerning earlier results in the theory of spectral asymptotics for differential and
related estimates for pseudodifferential operators, we refer to [2, 3, 14, 22].
Our approach used in [28 – 31] was based on the following two steps:
Step 1: regularization of irregular coefficients allowing to reduce the problem to

the analogical problem for ‘smooth operators’;
Step 2: the analysis of ‘smooth operators’ in the spirit of [6], i.e. reduction to a
study of a pseudodifferential operator P of degree 2mµ and a proof of spectral
asymptotics for P via the Tauberian reasoning based on the analysis of
Z
u(t, y) = e−it λ dλ e(P , y, y, λ) (1.4)
(the Fourier transform of the spectral function of P ).
Concerning the sharp spectral asymptotics, we begin by noting that Step 1 can
be made adopting the regularization procedure used in [28 – 31]. However, pro-
found differences between intermediate and sharp remainder estimates appear in
SHARP SPECTRAL ASYMPTOTICS AND WEYL FORMULA 293
Step 2. Indeed, the proof of [29] uses a pseudodifferential calculus based on the
fact that P is a pseudodifferential operator of degree 2mµ < 1 and the evolution
propagator e−it P does not change the wave front set. To obtain sharp remainder
estimates, we must use P of degree 1 and we expect that e−it P propagates the
wave front set along the trajectories of the Hamiltonian flow. Thus, a reasonable
regularity hypothesis should allow to solve and to control the Hamiltonian system.
A good hypothesis of this type is the hypothesis that the Hamiltonian field is Lip-
schitz continuous, i.e. the hypothesis that first-order derivatives of coefficients are
Lipschitz continuous.
We note that the analysis of all spectral problems mentioned before begins by
localization. More precisely, it suffices to study functions supported in a suitable
open subset of M and the corresponding quadratic form in local coordinates can be
considered on the whole Rd , i.e. it suffices to consider the quadratic form
X
A0 [ϕ, ψ] = 0
(aα,β D α ϕ, D β ψ) (1.5)
|α|,|β|6m
for ϕ, ψ ∈ C0∞ (Rd ), where aα,β

0
∈ L∞ (Rd ), aβ,α
0 0
= aα,β (we denoted by (·, ·) the
|α| α
scalar product of L (R ) and D = (−i) ∂ ).
2 d α
Then the ellipticity hypothesis of Theorem 1.1 means that there is c > 0 such
that
X
0 0
apr (x, ξ ) = aα,β (x)ξ α+β > c|ξ |2m (1.6)
|α|=|β|=m
and the regularity hypothesis of Theorem 1.1 means that there is C > 0 such that
0 0
|∇aα,β (x) − ∇aα,β (y)| 6 C|x − y| for x, y ∈ Rd , (1.7)
where it suffices to assume (1.7) for |α| = |β| = m.
Further on we fix 0 6 δ < 1 and recall that Step 1 of our approach leads
to pseudodifferential operators with symbols belonging to Hörmander’s classes of
type 1, δ (defined in Section 2 of this paper). More precisely, under the regular-
0
ity condition (1.7), instead of apr introduced in (1.6) we consider a symbol a 0
satisfying
0 0
|∂ξα ∂xα a 0 (x, ξ )| 6 Cα,α 0 hξ i2m−|α|+δ(|α |−2)+ (1.8)
for (x, ξ ) ∈ Rd × Rd , α, α 0 ∈ Nd , where hξ i = (1 + |ξ |2 )1/2 and s+ denotes
the positive part of the real number s, i.e. (|α 0 | − 2)+ = 0 when |α 0 | 6 2 and
(|α 0 | − 2)+ = |α 0 | − 2 when |α 0 | > 2.
Moreover, similarly as in [28] (using (2.6) of [30]) the ellipticity and homo-
0
geneity of apr guarantee the existence of constants C, c > 0 such that
|a 0 (x, ξ )| > c|ξ |2m − C, (1.9)
|∇ξ a 0 (x, ξ )| > c|ξ |2m−1 − C. (1.10)

294 LECH ZIELINSKI
For s ∈ R, let H s be the corresponding Sobolev space on Rd . We say that R is

smoothing and write R ∈ 9 −∞ if R is a linear operator on the Schwartz space
S(Rd ) possessing continuous extensions H −s → H s for every s ∈ R. The class
9 −∞ is a Fréchet space with seminorms ||R||B(H −n ,H n ) , n ∈ N, where B(X, X0 ) is
the Banach space of linear bounded operators X → X0 .
As explained in [28 – 30], asymptotic estimates for A0 follow from analogical
estimates for the pseudodifferential operator
A0 = a 0 (x, D) + R 0 with R 0 ∈ 9 −∞ , (1.11)
where (1.11) means that for ϕ ∈ S(Rd ) we define
Z
0 −d
(A ϕ)(x) = (2π ) eixξ a 0 (x, ξ )ϕ̂(ξ ) dξ + (R 0 ϕ)(x)
and ϕ̂ denotes the Fourier transform of ϕ.

Step 2 of our approach consists in the proof of
THEOREM 1.2. Let m > 0 and 0 6 δ < 1. Let A0 be given by (1.11) with
a 0 satisfying (1.8)–(1.10) and assume that A0 is self-adjoint in L2 (Rd ) (with the
domain H 2m ). Then A0 is bounded from below, every spectral projector E(A0 , λ) ∈
9 −∞ has a smooth integral kernel e(A0 , ·, ·, λ) and the estimate
e(A0 , y, y, λ) = ω(a 0 , y, λ)(1 + O(λ−1/(2m))) (1.12)
holds with
Z
0 −d
ω(a , y, λ) = (2π ) dξ,
Re a 0 (y,ξ )<λ
uniformly with respect to y ∈ Rd .

Following [29, Section 5], it is easy to see that Theorem 1.2 implies Theo-
rem 1.1.
The remaining part of this paper is devoted to the proof of Theorem 1.2 in the
special case δ < 1/2. The additional hypothesis δ < 1/2 allows us to simplify
the proof by applying directly an adequate theory of Fourier integral operators (cf.
Section 2). It is easy to see that the restriction δ < 1/2 does not allow us to obtain
the assertion of Theorem 1.1, but it does allows us to obtain the following weaker
result:
THEOREM 1.10 . Let M and A be as in Theorem 1.1. Then (1.2) holds with µ =
1/2m if second order derivatives of coefficients (in local coordinates) are Hölder
continuous with a certain exponent r > 0.
The proof of Theorem 1.2 in its general form will be given in the subsequent
paper [32], where we will develop the analysis described in this paper.
At the end of this introduction, we note that the regularization procedure of Step
1 was considered in the paper of Kumano-Go, Nagase [12] to study hyperbolic sys-
tems (cf. [8] and [24]). In particular, we follow their notations for classes Sρm ((r))
and we refer to the book of Kumano-Go [11, Chapter 10, Section 1] for the details
concerning the justification of the composition formulas for pseudodifferential and
Fourier integral operators described in Section 4. The properties of the Hamiltonian
systems considered in Lemma 4.1 are well known (cf. [8, 13, 24]) but we join their
simple proof in the Appendix. However, instead of the theory of multiproducts
of phase functions considered in [8, 13, 24], we preferred to present the standard
construction of Fourier integral operators via transport equations with an adequate
version of the Egorov theorem needed in the proof of Theorem 1.2.
2. Notations and Ideas of the Proof

We begin by recalling Hörmander’s classes of symbols of type ρ, δ (cf. [5]). If
0
m ∈ R, 0 6 δ < ρ 6 1, d, d 0 , d 00 ∈ N \ {0} and X is an open set of Rd , then
00
m
Sρ,δ (X × Rd ) denotes the class of functions a: X × Rd → Cd satisfying
0 0
|∂ξα ∂xα a(x, ξ )| 6 Cα,α 0 hξ im−ρ|α|+δ|α | (2.1)
0
for (x, ξ ) ∈ X × Rd , α ∈ Nd , α 0 ∈ Nd .
We say that {av }v∈V is a bounded subset of Sρ,δm
(X × Rd ) if a = av satisfies
(2.1) with constants Cα,α 0 independent of the parameter v ∈ V . We write
\
m
Sρ,δ = Sρ,δ
m
(Rd × Rd ), S −∞ (X × Rd ) = m
Sρ,δ (X × Rd )
m∈R
and we recall that a(x, D) ∈ 9 −∞ ⇔ a ∈ S −∞ .

We denote by 9ρ,δm
the class of operators of the form a(x, D)+R with R ∈ 9 −∞
and a ∈ Sρ,δ
m
. For 1/2 < ρ 6 1, we write
m
Sρ,1−ρ (X × Rd ) = Sρm (X × Rd ), m
Sρ,1−ρ = Sρm , m
9ρ,1−ρ = 9ρm .
The following preliminary result is proved in the Appendix:
LEMMA 2.1. Let A0 > I be as in Theorem 1.2 and let a 0 be its symbol satisfying
Re a 0 > 1. If P = A0 1/(2m), then P = p(x, D) + R with R ∈ 9 −∞ and
∂xα p ∈ S1,δ
1
if |α| 6 2, (2.2)
p − (Re a 0 )1/(2m) ∈ S1,δ

0
, (2.3)
|p(x, ξ )| > c|ξ | for |ξ | > C, (2.4)
|∇ξ p(x, ξ )| > c for |ξ | > C, (2.5)

296 LECH ZIELINSKI
where C, c > 0 are certain constants.
Similarly as in [6], the proof of Theorem 1.2 follows from the asymptotic
formula for the spectral function of P ,
e(P , y, y, λ) = ω(p0 , y, λ)(1 + O(λ−1 )), (2.6)
with
Z
−d
ω(p0 , y, λ) = (2π ) dξ,
p0 (y,ξ )<λ
where we have denoted p0 = Re p.

Since P is elliptic of degree 1, for every n ∈ N we have
Im t < 0 ⇒ ||e−it P ||B(H −n ;H n ) 6 CN ||P n e−it P P n || 6 CN |Im t|−2n
and introducing the holomorphic function u(·, y) in the lower complex half-plane
C− = {t ∈ C: Im t < 0} by
Z
u(t, y) = e−it λ dλ e(P , y, y, λ) = he−it P δy , δy i for t ∈ C− (2.7)
(where δy denotes the delta of Dirac in y) we can apply Theorem 3.1.11 of [7] to
define the distributional boundary value
u(t, y) = S 0 (R) − lim u(t − iε, y) for t ∈ R. (2.8)

ε&0
Then the approach of [6] uses the expression of e−it P by means of Fourier integral
operators to write
Z
u(t, y) = (2π )−d ei 8̃(t,y,ξ )q(t, y, ξ ) dξ for t ∈ ]−θ; θ[, (2.9)
where θ > 0 is small enough, and to show that u(·, y) ∈ C ∞ (]−θ; θ[\{0}) has a
singularity in 0 with the principal part
Z Z
−it λ −d
e dλ ω(p0 , y, λ) = (2π ) e−itp0 (y,ξ )d ξ.
In the case δ < 1/2, our proof of Theorem 1.2 will use the analogical schema. In
particular, we shall prove Theorem 2.4 allowing us to express e−it P as a Fourier
integral operator with an amplitude q ∈ S1−δ,δ
0
satisfying q(0, y, ξ ) = 1. However,
(2.6) cannot be obtained from the result of Theorem 2.4 in the way described in
[6]. Indeed, even if we could replace 8̃ by −tp0 in (2.9), the only estimate of the
time derivative of q we know is ∂t q ∈ S1−δ,δ
δ
(cf. the remark at the end of the
proof of Theorem 2.4 in Section 4), which is not sufficient to follow the Tauberian
reasoning of [6] based on the estimate q(t, y, ξ ) = 1 + O(t) when t → 0.
Thus, we propose another approach of studying u(·, y). Similarly as in [28, 29],
we begin by noting that e−it P is a bounded family of pseudodifferential operators
of type 1, δ if t belongs to an angular region {t ∈ C− : −Im t > θ 0 |Re t|} with
θ 0 > 0. In Section 3, we describe a parabolic-type approximation of e−it P by a
sequence of operators
QN (t) = (e−itp0 qN (t))(x, D), (2.10)
defining a sequence of holomorphic functions in C− by the formula
Z
−d
uN (t, y) = hQN (t)δy , δy i = (2π ) e−itp0 (y,ξ )qN (t, y, ξ ) dξ. (2.11)
Then the distributional boundary value for t ∈ R gives uN (·, y) smooth on R \ {0}
and our aim is to prove
THEOREM 2.2. Let θ > 0 be small enough. Then for every l0 ∈ N, we can find
N = N(l0 ) ∈ N and C = C(N, l0 ) > 0 such that
|∂tl (u − uN )(t, y)| 6 C (2.12)
holds for t ∈ ]−θ; θ[, y ∈ Rd and l = 0, . . . , l0 .
It seems interesting to underline the simplicity of the construction of uN based

on explicit differentiations and integrations giving, however, a full description of
u(·, y) modulo a smooth function. Even in the case δ = 0, this description seems
more explicit than the standard representation by Fourier integral operators where
the amplitude and the phase are solutions of some differential equations.
Another advantage of the parabolic construction is that all the time we can stay
within the framework of symbols of type 1, δ and integrating by parts we can find
a symbol qN0 = O(1) such that
Z
−d
uN (t, y) = (2π ) e−itp0 (y,ξ )(1 + tqN0 (t, y, ξ )) dξ, (2.13)
which allows us to obtain (2.6) from (2.12) via the Tauberian argument of [6].
An idea similar to integrations by parts is used in Section 5 to prove (2.12) and
the reasoning is based on a result being a version of the Egorov theorem, saying
that eit P a(x, D)e−it P is a pseudodifferential operator of type 1 − δ, δ if the symbol
a is of type 1 − δ, δ. A suitable version of the Egorov theorem is proved in Section
4 by means of Fourier integral operators.
Before giving more details about our framework of Fourier integral operators,
let us consider the change of variables
ã(y, η) = a(x(y, η), ξ(y, η)) (2.14)
defined by functions x: Y × Rd → X, ξ : Y × Rd → Rd , where X, Y are open sets
0
of Rd .
298 LECH ZIELINSKI
We consider the following decomposition of the associated Jacobi matrix

 ∂x ∂x 
!
(∂yk xj )↓j=1,...,d 0 (∂ηk xj )↓j=1,...,d 0
 ∂y ∂η  →k=1,...,d 0 →k=1,...,d
 =
∂ξ ∂ξ (∂yk ξj )↓j=1,...,d (∂ηk ξj )↓j=1,...,d
→k=1,...,d 0 →k=1,...,d
∂y ∂η
and assume
∂x ∂x
∈ Sρ0 (Y × Rd ), ∈ Sρ−1 (Y × Rd ), (2.15i)
∂y ∂η
∂ξ ∂ξ
∈ Sρ1 (Y × Rd ), ∈ Sρ0 (Y × Rd ), (2.15ii)
∂y ∂η
where a matrix b = (bj,k )↓j=1,...,d1 is considered as an element of Cd1 d2 with

→k=1,...,d2
X 1/2
|b| = |bj,k | 2
.
16j 6d1 , 16k 6d2
The following simple result is proved in the Appendix:
LEMMA 2.3. (a) Let a ∈ Sρm (X×Rd ) and let ã be defined by (2.14) with functions
x, ξ satisfying (2.15i,ii). Assume, moreover, that there is a constant C > 0 such that
C −1 hηi 6 hξ(y, η)i 6 Chηi (2.16)
holds for all (y, η) ∈ Y × Rd . Then ã ∈ Sρm (Y × Rd ).

(b) Assume, moreover, that
∂x a = (∂x1 a, . . . , ∂xd 0 a) ∈ Sρm (X × Rd ),

∂ξ a = (∂ξ1 a, . . . , ∂ξd a) ∈ Sρm−1 (X × Rd ).
Then
∂y ã = (∂y1 ã, . . . , ∂yd 0 ã) ∈ Sρm (Y × Rd ),

∂η ã = (∂η1 ã, . . . , ∂ηd ã) ∈ Sρm−1 (Y × Rd ).
NOTATIONS. Let m, d, d 0 , d 00 , X be as above, 1/2 < ρ 6 1 and r > 0. Then

00
Sρm ((r))(X × Rd ) denotes the class of functions a: X × Rd → Cd satisfying
0 0
|∂ξα ∂xα a(x, ξ )| 6 Cα,α 0 hξ im−|α|+(1−ρ)(|α|+|α |−r)+ (2.17)
0
for (x, ξ ) ∈ X × Rd , α ∈ Nd , α 0 ∈ Nd , where, as before, s+ denotes the positive
part of the real number s.
It is easy to check that
aj ∈ Sρ j ((r))(X × Rd ) for j = 1, 2 H⇒ a1 a2 ∈ Sρm1 +m2 ((r))(X × Rd )

m
and conditions (2.15i,ii) hold if x ∈ Sρ0 ((1))(Y × Rd ), ξ ∈ Sρ1 ((1))(Y × Rd ).

Further on, for θ > 0 we denote
Cθ = (]−θ; θ[×Rd ) × Rd . (2.18)
In Section 4 we give the proof of
THEOREM 2.4. Let P be as in Lemma 2.1. Assume that δ < 1/2 and set ρ =
1 − δ. If θ > 0 is small enough then
e−it P = Q(t) + R(t) with
Q(t) = (ei 8̃ q)(t, x, D), R ∈ C ∞ (]−θ; θ[; 9 −∞ ),
where
8̃ ∈ Sρ1 ((2))(Cθ ) is real valued, 8̃(0, ·, ·) = 0,
q ∈ Sρ0 (Cθ ), q(0, ·, ·) = 1.
Here Q(t) = (ei 8̃ q)(t, x, D) means that Q(t) is the operator

Z
−d
(Q(t)ϕ)(x) = (2π ) ei8(t,x,η) q(t, x, η)ϕ̂(η) dη, (2.19i)
where
8(t, x, η) = xη + 8̃(t, x, η) (2.19ii)
for ϕ ∈ S(Rd ). According to this notation, in the case 8̃ = 0 (identically), we

obtain the usual pseudodifferential operator (ei 8̃ q)(t, x, D) = q(t, x, D).
3. Parabolic Approximation
Let p be as in Lemma 2.1 and denote p0 = Re p. We note that (2.3) implies
p − p0 = Im p ∈ S1,δ
0
.
Let N ∈ N, N > 3 and define PN : C ∞ (Rdx × Rdξ ) → C ∞ (Rdx × Rdξ ) by
X
PN : a → PN a = (−i)|α| ∂ξα (a ∂xα p̄)/α!. (3.1)
|α|6N
300 LECH ZIELINSKI
Then we have
PROPOSITION 3.1. Set δ̄ = max{δ, 2/3}. Then there exist symbols

(k−1)δ̄
qN,1 ∈ S1,δ
0
, qN,2 ∈ S1,δ
1
, qN,k ∈ S1,δ for k > 3,
such that
(∂t + iPN )(qN (t)e−itp0 ) = q̃N0 (t)e−itp0

with (3.2)
X X
qN (t) = 1 + k
t qN,k , q̃N0 (t) = t k q̃N,k
0
, 0
q̃N,k ∈ k δ̄
S1,δ .
16k 6N N 6k 62N
P
Proof. We construct qn (t) = 06k6n t k qN,k by induction with respect to n =
0, 1, . . . , N, satisfying
X
P̃N qn (t) = t k q̃N,n,k
0
, (3.3(n))
n6k 6n+N
where
P̃N q(t) = eitp0 (∂t + iPN )(q(t)e−itp0 ).
As in [28], we note that
P̃N qn+1 (t) = t n ((n + 1)qN,n+1 + q̃N,n,n

0
)+
X
+ t n+1 P̃N qN,n+1 + t k q̃N,n,k
0
,
n+16k 6n+N
i.e. if (3.3(n)) holds, then taking
qN,n+1 = −q̃N,n,n
0
/(n + 1), (3.4)
we cancel the term with t n and (3.3(n + 1)) holds.

If q(t) does not depend on t, i.e. q(t) = q ∈ C ∞ (Rdx × Rdξ ), then
X
P̃N q(t) = t k q̃k
06k 6N
with
X
q̃0 = i(p̄ − p0 )q − (−i)|α|+1 ∂ξα (q ∂xα p̄)/α!,
16|α|6N
X
q̃k = cα0 ,...,αk ∂ξα0 (q ∂xα0 +···+αk p̄)∂ξα1 p0 . . . ∂ξαk p0 (1 6 k 6 N)
|α0 +···+αk |6N
αj 6=0 if j6=0
and it is easy to check that

X
q ∈ S1,δ
m
H⇒ P̃N q(t) ∈ S1,δ
m
+ tS1,δ
m+1
+ m+k δ̄
t k S1,δ , (3.5)
26k 6N
where we used the following notation:

X m(k)
X m(k)
a(t) ∈ t k S1,δ ⇐⇒ a(t) = t k ãk with ãk ∈ S1,δ .
k∈K k∈K
In particular q0 (t) = qN,0 = 1 identically implies

X
P̃N q0 (t) ∈ S1,δ
0
+ tS1,δ
1
+ t k S1,δ
k δ̄
. (3.6)
26k 6N
Using (3.4), we have qN,1 = −q̃N,0,00

∈ S1,δ
0
, hence
X
t P̃N qN,1 ∈ tS1,δ
0
+ t 2 S1,δ
1
+ t k+1 S1,δ
k δ̄
. (3.7)
26k 6N
Since 1 6 2δ̄, (3.6) and (3.7) imply

X
P̃N q1 (t) ∈ tS1,δ
1
+ t k S1,δ
k δ̄
. (3.8)
26k 6N+1
Next, (3.8) and (3.4) give qN,2 = −q̃N,1,10

/2 ∈ S1,δ1
, hence
X
t 2 P̃N qN,2 ∈ t 2 S1,δ
1
+ t 3 S1,δ
2
+ 1+k δ̄
t k+2 S1,δ . (3.9)
26k 6N
Since 2 6 3δ̄ and 1 + k δ̄ 6 (2 + k)δ̄, we obtain

X
P̃N qn (t) = t k S1,δ
k δ̄
(3.10(n))
n6k 6n+N
for n = 2.
Finally, if (3.10(n)) holds for a given n ∈ {2, . . . , N}, then we have
qN,n+1 = −q̃N,n,n
0
/(n + 1) ∈ S1,δ
nδ̄
and
X (n+k)δ̄
t n+1 P̃N qN,n+1 ∈ t n+1 S1,δ
nδ̄
+ t n+2 S1,δ
1+nδ̄
+ t n+1+k S1,δ ,
26k 6N
hence 1 + nδ̄ 6 (n + 2)δ̄ implies (3.10(n + 1)). 2
Further on, we consider y0 ∈ Rd , r > 0 small enough (independent of y0 ) and

show that the estimate (2.6) holds for y ∈ B(y0 , r) = {x ∈ Rd : |x − y0 | < r}.
302 LECH ZIELINSKI
Since all constants are translation invariant (can be chosen independent of y0 ), the
estimate (2.6) will hold for all y ∈ Rd .
For c > 0 and j = 1, . . . , d we introduce
0±j (c) = {ξ ∈ Rd : ±∂ξj p0 (y0 , ξ ) > 2c}. (3.11)
Due to (2.5), we can choose c > 0 small enough to have
[
0j (2c) ⊃ {ξ ∈ Rd : |ξ | > 1/c}. (3.12)
0<|j |6d
Moreover, we can choose r > 0 such that

(|x − y0 | < 4r and ξ ∈ 0±j (c/3)) H⇒ ±∂ξj p0 (x, ξ ) > c/3. (3.13)
The following lemma is proved in the Appendix:
0
LEMMA 3.2. (a) If c, r > 0, then there exist χj,c,r ∈ S1,0
0
such that
0
supp χj,c,r ⊂ B(y0 , 3r) × 0j (c/2) and 0
χj,c,r = 1 on B(y0 , 2r) × 0j (c).
(b) If c > 0 is small enough to guarantee (3.12), then there exist χj,c,r ∈ S1,0
0
(j = ±1, . . . , ±d) such that supp χj,c,r ⊂ B(y0 , 2r) × 0j (c) and
X
χj,c,r (x, ξ ) = 1, when x ∈ B(y0 , 3r/2), |ξ | > 1/c.
0<|j |6d
Similarly, as in [28, 29], we have
COROLLARY 3.3. In order to prove Theorem 1.2 it suffices to prove that (2.12)
holds if u is given by (2.7) and uN is given by (2.11) with qN described in Proposi-
tion 3.1.
Proof. It suffices to show the existence of qN0 ∈ S 0 (Cθ ) satisfying (2.13). If
χj,c,r are as in Lemma 3.2, then for k = 2, . . . , N there exists qN,k,j ∈ S1,δ k−1
such
that χj,c,r qN,k = (i∂ξj p0 ) qN,k,j and it remains to write
k−1
Z Z
e−itp0 (y,ξ )t k χj,c,r qN,k (y, ξ ) dξ = e−itp0 (y,ξ )t∂ξk−1
j
qN,k,j (y, ξ ) dξ. 2
00
Further, q ∈ Sρ,δ m
(X × Rd × Rd ) means that q(x, ξ, x 0 ) = q̃((x, x 0 ), ξ ) with
00
q̃ ∈ Sρ,δ
m
((X × Rd ) × Rd ).
Then for q ∈ Sρ,δ m
(Rd × Rd × Rd ) we write Op(qe−itp0 ) to denote the operator
on S(Rd ) with the distributional kernel
Z
0
Kt (q)(x, x 0 ) = (2π )−d ei(x−x )ξ −itp0 (x,ξ )q(x, ξ, x 0 ) dξ.
Let A∗ denote the adjoint of A. Then the calculus of [28, Section 2] gives
(∂t QN (t) + iQN (t)p(x, D)∗ ) = Op((q̃N0 + q̃N1 )(t)e−itp0 ), (3.14)
with q̃N0 given by (3.2) and

Z 1
q̃N1 (t)(x, ξ, x 0 ) =eitp0 (x,ξ )
(N + 1) ds(1 − s)N q̃N2 (t, s)(x, ξ, x 0 ),
0
where
q̃N2 (t, s)(x, ξ, x 0 )

X (−i)|α|
= ∂ξα ((qN (t)e−itp0 )(x, ξ )∂xα p̄(x + s(x 0 − x), ξ )).
|α|=N+1
α!
n+1−N(1−δ̄)
The form of qN (t) gives q̃N,n ∈ S1,δ (Rd × Rd × Rd ) such that
X
(q̃N0 + q̃N1 )(t) = t n q̃N,n , (3.15)
06n62N+1
and (3.14)–(3.15) give
[hQN (t − τ )eiτ P δy , δy i]ττ =Re

=0
t
(3.16)
Z Re t X
= dτ heiτ P δy , Op((t − τ )n q̃N,n e−i(t −τ )p0 )∗ δy i.
0 06n62N+1
However, for every θ, θ 0 > 0, n ∈ N, there is a constant Cn,θ,θ 0 such that
|(t − τ )n ei(τ −t )p0 (x,ξ )| 6 Cn,θ,θ 0 hξ i−n
holds when −Im t > θ 0 |Re t|, |t| < θ, τ ∈ [0; Re t].
It is easy to see (cf. [29, Section 5]) that the left-hand side of (3.16) tends to
(u − uN )(y, Re t) (in S 0 (Rd )) when Im t → 0, hence (3.15) implies the estimate
(2.12) for −Im t > θ 0 |Re t|, |t| < θ, τ ∈ [0; Re t], k = 0, N large enough. It
is clear that the analogical reasoning may be applied to estimate ∂tk (u − uN ) for
k = 0, . . . , k 0 when N = N(k 0 ) is large enough.
It remains to prove estimates (2.12) for t ∈ 40 (θ, θ 0 ), where
40 (θ, θ 0 ) = {t ∈ C : 0 < −Im t < θ 0 |Re t| and |t| < θ} (3.17)
and θ, θ 0 > 0 small enough.

Let us introduce V ⊂ Rd × C × R and Ṽ ⊂ V × R setting
V = {v = (y, t, τ 0 ) : y ∈ B(y0 , r), t ∈ 40 (θ, θ 0 ), τ 0 ∈ [0; Re t]}, (3.18)
Ṽ = {(v, τ ) : v = (y, t, τ 0 ) ∈ V and τ = τ 0 }. (3.19)

304 LECH ZIELINSKI
For 0 6 δ < ρ 6 1, n ∈ N, k ∈ Z, we introduce the conditions

m(n,k)
{qn,k,v }v∈V is a bounded subset of S1,δ , (3.20(n,k))
0
{an,k,v }v∈V is a bounded subset of Sρm (n,k) (Cθ ), (3.21(n,k)i)
for some m(n, k), m0 (n, k) ∈ R.

If a ∈ C ∞ (Cθ ), then a|τ =0 ∈ C ∞ (Rd × Rd ) denotes the function a|τ =0 (x, ξ ) =
a(0, x, ξ ). Besides (3.21(n,k)i) we consider the condition
0
m (n,k)+l(1−ρ)
{∂τl an,k,v |τ =0 }v∈V is a bounded subset of S1,1−ρ (3.21(n,k)ii)
for every l ∈ N.
For qn,k,v satisfying (3.20(n,k)), an,k,v satisfying (3.21(n,k)i,ii), we denote
J (qn,k,v , an,k,v )(τ ) = he−iτ P δy , an,k,v (τ, x, D) Op(qn,k,v ei(τ −t )p0 )∗ δy i,

]
where
v = (y, t, τ 0 ) ∈ V qn,k,v (x, ξ, x 0 ) = qn,k,v (x 0 , ξ ).
]
and
In Section 5 we shall prove the following:
PROPOSITION 3.4. Let n ∈ N \ {0}, let P be as in Lemma 2.1 and Ṽ as above.

We assume that δ < 1/2 and set ρ = 1 − δ. If qn,0,v satisfy (3.20(n,0)) and an,0,v
satisfy (3.21(n,0)i,ii), then there exist k0 = k0 (n) ∈ N, qn,k,v satisfying (3.20(n,k)),
an,k,v satisfying (3.21(n,k)i,ii) for k = 1, . . . , k0 with
m(n, k) + m0 (n, k) 6 m(n, 0) + m0 (n, 0) − n
and there is a constant C > 0 such that for (v, τ ) ∈ Ṽ , one has

n X
t J (qn,0,v , an,0,v )(τ ) − J (qn,k,v , an,k,v )(τ ) 6 C. (3.22(n))

16k 6k0
To end this section we check that Proposition 3.4 implies Theorem 2.2. Let q̃N,n
be given by (3.15). Then setting
qn,0,v (x, ξ ) = (1 − τ 0 /t) q̃N,n (y, ξ, x) for v = (y, t, τ 0 ) ∈ V
n
we find that (3.20(n,0)) holds with m(n, 0) = n + 1 − N(1 − δ̄) and

Op((t − τ )n q̃N,n ei(τ −t )p0 )∗ δy = Op(t n qn,0,v ei(τ −t )p0 )∗ δy
]
for (v, τ ) ∈ Ṽ allows to write the right-hand side of (3.16) in the form
Z X
dτ t n J (qn,0,v , an,0,v )(τ ), (3.23)
[0; Re t ] 06n62N+1
where an,0,v (τ, x, ξ ) = 1 identically.

If N is large enough to ensure 1 − N(1 − δ̄) 6 −2(d + 1), then Proposition
3.4 allows us to write (3.22(n)) with qn,k,v satisfying (3.20(n,k)), an,k,v satisfying
(3.21(n,k)i,ii) and m(n, k) + m0 (n, k) 6 −2(d + 1) for k > 0, i.e.

J (qn,k,v , an,k,v )(τ ) = O(1) for k > 1 H⇒ t n J (qn,0,v , an,0,v )(τ ) = O(1)
uniformly with respect to (v, τ ) ∈ Ṽ (cf. the last remark of Appendix) and u −
uN = O(1) follows from (3.23).
Calcultaing the lth derivative of (3.23), we obtain a linear combination of sim-
ilar expressions, where an,0,v (τ, ·, ·) is a symbol of a power of P (i.e. satisfying
(3.21(n,0)i,ii) with m0 (n, 0) 6 l) and qn,0,v satisfy (3.20(n,0)) with m(n, 0) 6 l +
n+1−N(1− δ̄). Therefore, the general statement of Theorem 2.2 holds by a similar
reasoning if N = N(l0 ) is large enough to ensure 2l0 + 1 − N(1 − δ̄) 6 −2(d + 1).
4. Fourier Integral Operators

Let p0 ∈ Sρ1 ((2)) be real-valued and let ϑ: R × Rd × Rd → Rd × Rd denote the
associated Hamitonian flow
ϑ(t, y, η) = exp(tHp0 )(y, η) = (x(t, y, η), ξ(t, y, η)), (4.1)
i.e.
∂t x(t, y, η) = ∂ξ p0 (ϑ(t, y, η)), ∂t ξ(t, y, η) = −∂x p0 (ϑ(t, y, η))
and ϑ(0, y, η) = (y, η). We write
x(t, y, η) = y + x̃(t, y, η), ξ(t, y, η) = η + ξ̃ (t, y, η). (4.2)
Due to |∂ξ p0 | 6 C and |∂x p0 (x, ξ )| 6 Chξ i, the Gronwall estimate ensures
|x̃(t, y, η)| 6 C|t| < 1/2, |ξ̃ (t, y, η)| 6 C|t|hηi 6 hηi/2 for |t| < θ
if θ > 0 is chosen small enough. Therefore
hηi/2 6 hξ(t, y, η)i 6 2hηi for |t| < θ (4.3)
and in the Appendix we check that
LEMMA 4.1. (a) If θ > 0 is small enough, then

x̃, ∂t x ∈ Sρ0 ((1))(Cθ ) and ξ̃ , ∂t ξ ∈ Sρ1 ((1))(Cθ ).
(b) If (t, η) ∈ ]−θ; θ[×Rd then the application y → x(t, y, η) is a bijection

of Rd and let x → y(t, x, η) be its inverse. Then
y(t, x, η) = x + ỹ(t, x, η) with ỹ ∈ Sρ0 ((1))(Cθ ).
306 LECH ZIELINSKI
Let y be as in Lemma 4.1. Then the standard construction of the associated

generating (action) function 8 ∈ C ∞ (Cθ ) gives
y = ∂η 8 and ξ(t, y, η) = ∂x 8(t, x(t, y, η), η), (4.4)
∂t 8(t, x, η) + p0 (x, ∂x 8(t, x, η)) = 0, 8(0, x, η) = xη. (4.5)
Introducing 8̃ given by (2.19ii) we have 8̃ ∈ Sρ1 ((2))(Cθ ). Indeed, ∂η 8̃ = ỹ ∈

Sρ0 ((1))(Cθ ) due to Lemma 4.1 and
∂x 8̃(t, x, η) = ξ̃ (t, y(t, x, η), η) = t ξ̄ (t, y(t, x, η), η) (4.6)

with
Z 1
ξ̄ (t, y, η) = ds(∂t ξ̃)(st, y, η),
0
hence ξ̄ ∈ Sρ1 ((1))(Cθ ) and Lemma 2.3 implies ∂x 8̃ ∈ Sρ1 ((1))(Cθ ) and ∂t 8̃ ∈
Sρ1 ((1))(Cθ ) by (4.5). Moreover, calculating the ∂η -derivative of (4.6) we find
|∂η ∂x 8̃(t, x, η)| 6 C|t|, hence choosing θ > 0 small enough, we have
|∂x 8̃(t, x, η)| 6 C|t||η| 6 |η|/2 for |t| < θ. (4.7)

Let ζ̃ ∈ Sρ1 ((1))((]−θ; θ[×Rd × Rd ) × Rd ) be given by
Z 1
0
ζ̃ (t, x, x , η) = ds(∂x 8̃)(t, x + s(x 0 − x), η). (4.8)
0
Due to (4.7), setting ζ(t, x, x 0 , η) = η + ζ̃ (t, x, x 0 , η) we have

hηi/2 6 hζ(t, x, x 0 , η)i 6 2hηi for |t| < θ. (4.9)
Writing 8(t, x, η) − 8(t, x 0 , η) = (x − x 0 )ζ(t, x, x 0 , η), we compute
a(t, x, D)(ei 8̃ q)(t, x, D) = (ei 8̃ q̃)(t, x, D), (4.10)

0
where q ∈ Sρm (Cθ ), a ∈ Sρm (Cθ ) and
Z
0 0
q̃(t, x, η) = (2π )−d ei(x−x )(ξ −ζ(t,x,x ,η)) a(t, x, ξ )q(t, x 0 , η) dx 0 dξ
Z
0 0
= (2π )−d ei(x−x )η q̃0 (t, x, η, x 0 , η0 ) dx 0 dη0 , (4.11)
with
q̃0 (t, x, η, x 0 , η0 ) = a(t, x, η0 + ζ(t, x, x 0 , η))q(t, x 0 , η).
Then modulo S −∞ (Cθ ) we may replace q̃0 by q̃0 χ with χ(η, η0 ) = χ0 (η0 hηi−1 ),
χ0 (x) = 1 for |x| 6 1/4, χ0 (x) = 0 for |x| > 1/3 (the contribution of the
0 0
integral (4.11) in the region |η0 | > hηi/4 is easy to estimate using |η0 |2n ei(x−x )η =
0 0
1nx0 ei(x−x )η to integrate by parts, cf. [11]). Therefore, the standard Taylor formula
0
in η = 0 gives the asymptotic expansion
X
q̃(t, x, η) ∼
= (−i)−|α| ∂xα0 ∂ηα0 q̃0 (t, x, η, x 0 , η0 )|x 0 =x, η0 =0 /α!.
α∈Nd
Using ζ(t, x, x, η) = ∂x 8(t, x, η), (4.9) and Lemma 2.3, we find

0 0
q̃ ∈ Sρm+m (Cθ ) and r̃0 = q̃ − ã0 q ∈ Sρm+m +1−2ρ (Cθ )
with
ã0 (t, x, η) = a(t, x, ∂x 8(t, x, η)). (4.12)
Since ∂xj ζ(t, x, x, η) = ∂xj ∂x 8(t, x, η)/2, setting
ã1,j (t, x, η) = (∂ξj a)(t, x, ∂x 8(t, x, η)), (4.13(j))
X
ã1 (t, x, η) = (∂ξα a)(t, x, ∂x 8(t, x, η))∂xα 8(t, x, η)/2, (4.14)
|α|=2
X
r̃1 = q̃ − ã0 q − i ã1 + ã1,j ∂xj q
16j 6d
and using Lemma 2.3, we find that for j = 1, . . . , d,

a ∈ Sρm ((1))(Cθ ) ⇒ ã1 ∈ Sρm−1 (Cθ ),
0
ã1,j ∈ Sρm−ρ (Cθ ), r̃0 , r̃1 ∈ Sρm+m −ρ (Cθ ), (4.15)
a ∈ Sρ1 ((2))(Cθ ) ⇒ ã1 ∈ Sρ0 (Cθ ),
0
ã1,j ∈ Sρ0 ((1))(Cθ ), r̃1 ∈ Sρm +1−2ρ (Cθ ). (4.16)
Proof of Theorem 2.4. Let ã1,j ∈ Sρ0 ((1))(Cθ ), ã1 ∈ Sρ0 (Cθ ), be obtained from
(4.13(j)) and (4.14) by using a(t, x, ξ ) = p0 (x, ξ ).
Let us introduce the first-order differential operator
X
L = ã1,0 + ∂t + ã1,j ∂xj , (4.17)
16j 6d
where ã1,0 = ã1 − i(p − p0 ) ∈ Sρ0 (Cθ ).

Since 8 satisfies the eiconal equation (4.5), using (4.10)–(4.15) we find
(i∂t + p(x, D))(ei 8̃ qn )(t, x, D) = (ei 8̃ (iLqn + F (qn )))(t, x, D),

where (4.18)
qn ∈ Sρm(n) (Cθ ) ⇒ F (qn ) ∈ Sρm(n)+1−2ρ (Cθ ).
308 LECH ZIELINSKI
P
Following the standard parametrix construction of q ∼ = n>0 qn , we take qn satis-
fying transport equations Lq0 = 0, q0 |t =0 = 1 and iLqn +F (qn−1 ) = 0, qn |t =0 = 0
for n = 1, 2, . . . . To ensure qn ∈ Sρm(n) (Cθ ) with m(n) = n(1 − 2ρ), it remains to
prove the following assertion:
Lqn = fn ∈ Sρm(n) (Cθ ), qn |t =0 = δ0,n H⇒ qn ∈ Sρm(n) (Cθ ), (4.19)
where δ0,0 = 1 and δ0,n = 0 for n > 1.
However, following the proof of Lemma 4.1, we can solve the system
∂t xj0 (t, y, η) = ã1,j (t, x 0 (t, y, η), η) (j = 1, . . . , d), x 0 (0, y, η) = y
with (4.20)
xj0 (t, y, η) = yj + x̃j0 (t, y, η), x̃j0 ∈ Sρ0 ((1))(Cθ ).
Introducing
0
a1,j (t, y, η) = ã1,j (t, x 0 (t, y, η), η), fn0 (t, y, η) = fn (t, x 0 (t, y, η), η),
Rt
Z t Rt
−
e−
0 (s,y,η) ds 0 (s,y,η)ds
qn0 (t, y, η) = δ0,n e 0 a1,0
+ s0 a1,j
fn0 (s 0 , y, η) ds 0
0
for j = 0, 1, . . . , d and using Lemma 2.3 we find that for j = 1, . . . , d,

0
a1,j ∈ Sρ0 ((1))(Cθ ), 0
a1,0 ∈ Sρ0 (Cθ ), fn0 ∈ Sρm(n) (Cθ ),
qn0 ∈ Sρm(n) (Cθ ), 0

(a1,0 + ∂t )qn0 = fn0 , qn0 |t =0 = δ0,n
and qn0 (t, y, η) = qn (t, x 0 (t, y, η), η). To complete the proof we note that choosing
θ > 0 small enough and reasoning as in the proof of Lemma 4.1, we find that the
application y → x 0 (t, y, η) is a bijection of Rd and its inverse
x → y 0 (t, x, η) = x + ỹ 0 (t, x, η) with ỹ 0 ∈ Sρ0 ((1))(Cθ ),
hence, Lemma 2.3 ensures qn ∈ Sρm(n) (Cθ ). 2
Remark. The above reasoning ensures ∂t q00 ∈ Sρ0 (Cθ ), but we do not know if
∂t q0 ∈ Sρ0 (Cθ ).
Since |∂ ζ̃ /∂η(t, x, x 0 , η)| 6 C|t|, following the proof of Lemma 2.3, we find
that η → ζ(t, x, x 0 , η) is a bijection of Rd with the inverse ζ → η(t, x, x 0 , ζ ) =
ζ + η̃(t, x, x 0 , ζ ) such that η̃ ∈ Sρ1 ((1))((]−θ; θ[×Rd × Rd ) × Rd ) and
hζ i/2 6 hη(t, x, x 0 , ζ )i 6 2hζ i. (4.90 )
m
If qj ∈ Sρ j (Cθ ) for j = 1, 2, then the operator
(ei 8̃ q1 )(t, x, D)(ei 8̃ q2 )(t, x, D)∗ = Op(q(t)) (4.21)

with
−1
∂ζ
0
q(t, x, ζ, x ) = q1 (t, x, η)q̄2 (t, x, η) det 0
(t, x, x , η) .
∂η
η=η(t,x,x 0,ζ )
is pseudodifferential with q ∈ Sρm1 +m2 (Cθ ) due to Lemma 2.3. Indeed, the distrib-
utional kernel of (4.21) is
Z
0 0
ei(x−x )ζ(t,x,x ,η) q1 (t, x, η)q̄2 (t, x, η) dη (4.22)
and it remains to make the change of variables ζ = ζ(t, x, x 0 , η) in the integral

(4.22).
Writing 8(t, x, η) − 8(t, x, η0 ) = (η − η0 )z(t, x, η, η0 ) with
Z 1
0
z(t, x, η, η ) = ds(∂η 8)(t, x, η + s(η0 − η)), (4.23)
0
0
for q ∈ Sρm (Cθ ), b ∈ Sρm (Cθ ) we compute
(ei 8̃ q)(t, x, D)b(t, x, D) = (ei 8̃ q̃)(t, x, D) (4.24)
with
Z
−d 0 0
q̃(t, x, η) = (2π ) ei(η−η )(y−z(t,x,η,η )) q(t, x, η0 )b(t, y, η) dy dη0
Z
−d 0 0
= (2π ) ei(η−η )x q̃0 (t, x, η, x 0 , η0 ) dx 0 dη0 , (4.25)
where
q̃0 (t, x, η, x 0 , η0 ) = q(t, x, η0 )b(t, x 0 + z(t, x, η, η0 ), η).
Similarly as before suitable integrations by parts show that modulo S −∞ (Cθ ) we

may replace q̃0 by q̃0 χ̃ with χ̃ (η, η0 ) = χ0 ((η − η0 )hηi−1 ) and χ0 as before. Then
the Taylor formula in x 0 = 0 gives the asymptotic expansion
X
q̃(t, x, η) ∼
= (−i)−|α| ∂xα0 ∂ηα0 q̃0 (t, x, η, x 0 , η0 )|x 0 =0, η0 =η /α!,
α∈Nd
0
i.e. q̃ ∈ Sρm+m (Cθ ) and since z(t, x, η, η) = ∂η 8(t, x, η), we have
0
q̃ − b̃0 q ∈ Sρm+m +1−2ρ (Cθ ) with b̃0 (t, x, η) = b(t, ∂η 8(t, x, η), η). (4.26)
COROLLARY 4.2. Let P be as in Lemma 2.1. Assume that δ < 1/2, ρ = 1 − δ

and a ∈ Sρm (Cθ ). Then there exists b0 ∈ Sρm (Cθ ) such that
eit P a(t, x, D)e−it P = b0 (t, x, D) + R(t), (4.27)

310 LECH ZIELINSKI
with R ∈ C ∞ (]−θ; θ[; 9 −∞ ),

b − b0 ∈ Sρm+1−2ρ (Cθ ) if b(t, y, η) = a(t, ϑ(t, y, η)). (4.28)
Proof. Due to Theorem 2.4, (4.10) and (4.24), we have
a(t, x, D)e−it P = (ei 8̃ (ã0 q + r̃0 ))(t, x, D) + R1 (t),
e−it P b(t, x, D) = (ei 8̃ (b̃0 q + r0 ))(t, x, D) + R2 (t),

with r0 , r̃0 ∈ Sρm+1−2ρ (Cθ ), ã0 given by (4.12), b̃0 by (4.26) and R1 , R2 ∈
C ∞ (]−θ; θ[; 9 −∞ ). If x = x(t, y, η) then
(x, ∂x 8(t, x, η)) = (x, ξ(t, y, η)) = ϑ(t, y, η) = ϑ(t, ∂η 8(t, x, η), η)
implies ã0 = b̃0 and we complete the proof using (4.21) in

eit P a(t, x, D)e−it P − b(t, x, D)
= (ei 8̃ q)(t, x, D)∗ (ei 8̃ (r0 + r̃0 ))(t, x, D) + R(t). 2
5. Proof of Theorem 1.2 in the Case δ < 1/2

At the end of Section 3 we noticed that it remains to prove Proposition 3.4. How-
ever, it suffices to prove the assertion of Proposition 3.4 for n = 1. Indeed, if Propo-
sition 3.4 holds for a given n ∈ N, then using tqn+1,k,v , an+1,k,v instead of qn,0,v ,
an,0,v , we can express t n+1 J (qn+1,0,v , an+1,0,v ) as a sum of tJ (qn+1,k,v , an+1,k,v )
and the assertion for n + 1 follows from (3.22(1)) with qn+1,k,v , an+1,k,v instead of
q1,0,v , a1,0,v .
Further on we treat Proposition 3.4 in the case n = 1 only and we prove
PROPOSITION 5.1. Assume that Ṽ is given by (3.19), q1,0,v satisfy (3.20(1,0)),

a1,0,v satisfy (3.21(1,0)i,ii). Then there exist k0 ∈ N, q1,k,v satisfying (3.20(1,k)),
a1,k,v satisfying (3.21(1,k)i,ii) for k = ±1, . . . , ±k0 with
m(1, k) + m0 (1, k) 6 m(1, 0) + m0 (1, 0) − 1 for k > 0, (5.1)
m(1, k) + m0 (1, k) 6 m(1, 0) + m0 (1, 0) + 1 − 2ρ for k < 0 (5.2)
and there is a constant C > 0 such that for (v, τ ) ∈ Ṽ one has

X
tJ (q1,0,v , a1,0,v )(τ ) − (J (q1,k,v , a1,k,v ) +

16k 6k0

+ tJ (q1,−k,v , a1,−k,v ))(τ ) 6 C. (5.3)
To justify that Proposition 5.1 implies Proposition 3.4 with n = 1, we note that
after having written (5.3) we may express J (q1,−k,v , a1,−k,v ) (k = 1, . . . , k0 ) in
the analogical way, i.e. applying Proposition 5.1 to q1,−k,v , a1,−k,v (k = 1, . . . , k0 )
instead of q1,0,v , a1,0,v . Repeating this procedure l times, we obtain the expression
(5.3) with some q1,k,v satisfying (3.20(1,k)), a1,k,v satisfying (3.21(1,k)i,ii) and
m(1, k) + m0 (1, k) 6 m(1, 0) + m0 (1, 0) − l(2ρ − 1) for k < 0.
Thus for l large enough we obtain the assertion of Proposition 3.4 with n = 1.
Further on, instead of writing the following double condition
{av }v∈V is a bounded subset of Sρm (Cθ ) and
m+l(1−ρ)
{∂τl av |τ =0 }v∈V is a bounded subset of S1,1−ρ for every l ∈ N
we write simply that {av }v∈V is a bounded subset of S̃ρm (Cθ ).

Then we have the following simple property
LEMMA 5.2. Let j ∈ {1, . . . , d} and let {av }v∈V be a bounded subset of S̃ρm (Cθ ).
If xj stands for the operator of multiplication by j th coordinate, then
[xj , av (τ, x, D)] = av+ (τ, x, D) + τ av− (τ, x, D), (5.4)
where
{av+ }v∈V is a bounded subset of S̃ρm−1 (Cθ ),
{av− }v∈V is a bounded subset of S̃ρm+1−2ρ (Cθ ).
Proof. We have
av (τ, x, ξ ) = av (0, x, ξ ) + τ bv (τ, x, ξ ),
with
Z 1
bv (τ, x, ξ ) = ds(∂τ av )(sτ, x, ξ ).
0
Since {bv }v∈V is a bounded subset of S̃ρm+1−ρ (Cθ ), it is easy to check that the
assertion of lemma holds if we take
av+ (τ, x, ξ ) = i∂ξj av (0, x, ξ ), av− (τ, x, ξ ) = i∂ξj bv (τ, x, ξ ). 2
Let us consider the operator

Z 1
P̃1 (τ ) = ds e−isτ P [iP , xj ]eisτ P = p̃1 (τ, x, ξ ) + R(τ ) (5.5)
0
with R ∈ C ∞ (]−θ; θ[, 9 −∞ ). Then

l(1−ρ)
p̃1 ∈ Sρ0 (Cθ ) and ∂τl p̃1 |τ =0 ∈ S1,1−ρ for every l ∈ N. (5.6)
312 LECH ZIELINSKI
Indeed, the first assertion (5.6) follows from Corollary 4.2 and to obtain the second
one we note that
∂τl P̃1 (0) = adl+1

iP xj = ∂τ p̃1 (0, x, D) + ∂τ R(0),
l l
where we have used the notation
ad1A B = adA B = [A, B], A B = [A, adA B]

adl+1 l
(l ∈ N)
m+1−ρ
and since A ∈ 91,1−ρ
m
⇒ [P , A] ∈ 91,1−ρ , it is clear that the induction with
l(1−ρ)
respect to l ∈ N gives adiP [P , xj ] ∈ 91,1−ρ for every l ∈ N.
l
Proof of Proposition 5.1. To begin, we note that due to Remark (a) at the end of
Appendix, the partition of unity on B(y0 , 3r/2) × Rd introduced in Lemma 3.2(b)
allows to replace q1,0,v by q1,0,v χj,c,r , where j = ±1, . . . , ±d. Further on, we
assume that j is fixed and j > 0 (to treat the case j < 0 it suffices to replace xj
and ∂ξj by x−j and −∂ξ−j in the reasoning described below). Thus we may assume
that
supp q1,0,v ⊂ B(y0 , 2r) × 0j (c). (5.7)
Further on l = 0, 1, . . . and Rl,v,τ are uniformly smoothing operators (with respect

to (v, τ ) ∈ Ṽ ), i.e. {Rl,v,τ }(v,τ )∈Ṽ are bounded subsets of 9 −∞ . In particular, taking
0
χj,c,r as in Lemma 3.2, we can introduce uniformly smoothing operators
)(x, D) Op(q1,0,v ei(t −τ )p0 )∗ for v = (y, t, τ 0 ) ∈ V .

]
(1 − χj,c,r
0
(5.8)
Indeed, using remarks and notations explained at the end of the Appendix we find
) • q1,0,v form a bounded subset of S −∞ due to (5.7).
]
that (1 − χj,c,r
0
For y ∈ Rd we introduce the operator
Py = py (x, D) with py (x, ξ ) = ∂ξj p0 (y, ξ )
(i.e. the symbol py ∈ S10 does not depend on the x-variable) and for v = (y, t, τ 0 ) ∈
V we set

τ0 τ0
P̃v (τ ) = P̃1 (τ ) + 1 − Py = p̃v (τ, x, D) + R0,v,τ . (5.9)
t¯ t¯
Here P̃1 (τ ) is given by (5.5), hence Corollary 4.2 ensures

0Z 1
τ τ0
ds∂ξj p0 (ϑ(sτ, y, η)) + 1 − ∂ξj p0 (y, η) − p̃v (τ, y, η)
t¯ t¯
0
6 Chηi1−2ρ
and due to (5.6) it is clear that {p̃v }v∈V is a bounded subset of S̃ρ0 (Cθ ).
Assuming that c, θ, θ 0 > 0 are small enough and τ ∈ ]−θ; θ[, y ∈ B(y0 , 3r),
η ∈ 0j (c/2), |η| > 1/c, we have ∂ξj p0 (ϑ(τ, y, η)) > c/3, hence there is χ0 ∈
C0∞ (Rd ) such that the symbols
p̃˜ v (τ, x, ξ ) = p̃v (τ, x, ξ )χj,c,r

0
(x, ξ ) + (1 − χj,c,r
0
(x, ξ )) + χ0 (ξ )
satisfy
|p̃˜ v (τ, x, ξ )| > c/4 for v ∈ V , (τ, x, ξ ) ∈ Cθ .
Setting av = a1,0,v /p̃˜ v it is easy to check that {av }v∈V is a bounded subset of
0
S̃ρm (1,0)(Cθ ) and we can decompose
a1,0,v (τ, x, D) = av (τ, x, D)p̃˜ v (τ, x, D) + a1,−1,v (τ, x, D) + R1,v,τ , (5.10)

where a1,−1,v satisfy (3.21(1, −1)i,ii) with m0 (1, −1) = m0 (1, 0) + 1 − 2ρ.
Due to (5.8) we have
av (τ, x, D)p̃˜ v (τ, x, D) Op(q0,1,v
]
ei(t −τ )p0 )∗
= av (τ, x, D)P˜v (τ ) Op(q0,1,v
]
ei(t −τ )p0 )∗ + R2,v,τ . (5.11)
Further on, we assume that τ = τ 0 . Introducing bv by the relation
[av (τ, x, D), P˜1 (τ )] = bv (τ, x, D) + R3,v,τ ,
and using (5.4) we can wirte
t¯ av (τ, x, D)P˜v (τ )
= av (τ, x, D)(xj − yj + (t¯ − τ 0 )Py ) +
+ (τ 0 P˜1 (τ ) − (xj − yj ))av (τ, x, D) +
+ (a1,2,v + t¯a1,−2,v )(τ, x, D) + R4,v,τ (5.12)
with

τ0 − τ0
a1,2,v = av+ , a1,−2,v = av + 1 − bv ,
t¯ t¯
i.e. a1,±2,v satisfy (3.21(1,±2)i,ii) with
m0 (1, 2) = m0 (1, 0) − 1 and m0 (1, −2) = m0 (1, 0) + 1 − 2ρ.
Moreover, the definition of P̃1 (τ ) gives
(xj − yj − τ P̃1 (τ ))e−iτ P δy = e−iτ P (xj − yj )δy = 0. (5.13)
Then the composition formula from the end of Appendix gives
]
Py Op(q1,0,v ei(τ −t )p0 )∗ = Op((py q1,0,v
]
+ qv )ei(τ −t )p0 )∗ + R5,v,τ , (5.14)
314 LECH ZIELINSKI
and introducing q1,−3,v (x, ξ ) = (1 − (τ 0 /t¯))qv (y, ξ, x), we note that
(t¯ − τ ) Op(qv ei(τ −t )p0 )∗ = t¯ Op(q1,−3,v

]
ei(τ −t )p0 )∗
and (3.20(1, −3)) holds with m(1, −3) = m(1, 0) + 1 − 2ρ.

Finally the integration by parts gives
(t¯ − τ )(Op(py q1,0,v ei(τ −t )p0 )∗ δy )(x)

]
(5.15)
Z
−d
ei(x−y)ξ (−i)∂ξj (ei(t¯−τ )p0 )q1,0,v (x, ξ ) dξ
]
= (2π )
]
= −(xj − yj )(Op(q1,0,v ei(τ −t )p0 )∗ δy )(x) + (Op(q1,3,v
]
ei(τ −t )p0 )∗ δy )(x),
] ]
where q1,3,v = −i∂ξj q1,0,v satisfy (3.20(1, 3)) with m(1, 3) = m(1, 0) − 1.
To complete the proof, we note that the choice of q1,k,v for k = ±3 as above,
a1,k,v for k = −1, ±2 as above, a1,1,v = 0 identically and
q1,k,v = q1,0,v , m(1, k) = m(1, 0) for k = ±1, ±2,

a1,k,v = a1,0,v , m0 (1, k) = m0 (1, 0) for k = ±3,
gives (5.3) due to the equalities (5.10)–(5.15). 2
6. Appendix
Proof of Lemma 2.1. If A0 ∈ 91,δ 2m

is elliptic of degree 2m, then it is well known that
0 1/(2m)
P =A ∈ 91,δ is elliptic of degree 1 (cf., e.g., [11]), hence Pj = [Dj , P ] ∈
1
m0 m00
91,δ , Pj,j 0 = [Dj 0 , Pj ] ∈ 91,δ with m0 6 1 + δ, m00 6 1 + 2δ. First we show that
A0 j = [Dj , A0 ] ∈ 91,δ 2m
⇒ m0 6 1. Indeed,
X
A0 j = [Dj , P 2m ] = P k Pj P 2m−1−k = 2mPj P 2m−1 + Bj (A.1)
06k 62m−1
and
0 k+m +δ−1 0 2m−1+m +δ−1 0
Pj ∈ 91,δ
m
⇒ [Pj , P k ] ∈ 91,δ ⇒ Bj ∈ 91,δ .
Therefore
max{1, m0 +δ−1}
2mPj = (A0 j − Bj )P 1−2m ∈ 91,δ
and we can take m0 6 max{1, m0 + δ − 1}, i.e. m0 6 1.

m00
Thus, we have proved ∂xα p ∈ S1,δ 1
for |α| 6 1 and assuming ∂xα p ∈ S1,δ for
|α| = 2 with m00 > 1, we find that the asymptotic expansion of the symbol of
m00 m00 +n−1
[Pj , P ] gives [Pj , P ] ∈ 91,δ . Next it is easy to see that [Pj , P n ] ∈ 91,δ for
m +2m−2 00
n ∈ N, hence (A.1) holds with Bj ∈ 91,δ . Therefore [Dj 0 , Bj ]P 1−2m ∈
00
m −1+δ m00 −1
91,δ and [Dj 0 , P n ] ∈ 91,δ
n
for n ∈ Z implies Bj [Dj 0 , P 1−2m ] ∈ 91,δ . Then
0 0 0
using the expression (A.1) and A j,j = [Dj 0 , A j ] ∈ 91,δ , we find that the operator
2m
2m[Dj 0 , Pj ] = [Dj 0 , A0 j − Bj ]P 1−2m + (A0 j − Bj )[Dj 0 , P 1−2m ]

max{1, m00 +δ−1}
belongs to 91,δ . Taking m00 6 max{1, m00 + δ − 1} we find m00 6 1.
Let p1 = (Re a 0 )1/(2m). Then (2.2)–(2.5) hold with p1 instead of p and since
[p1 (x, D)k , P n ] ∈ 91,δ
k+n−1
holds for n ∈ N, we obtain
(p1 (x, D)2m − P 2m ) − (p1 (x, D) − P )B ∈ 91,δ

2m−1
with
X
B= p1 (x, D)k P 2m−1−k .
06k 62m−1
Since the asymptotic formula for the symbol of a 0 (x, D)∗ gives Im a 0 ∈ S1,δ
2m−1
, we
have
p1 (x, D)2m − (Re a 0 )(x, D) ∈ 91,δ
2m−1
⇒ p1 (x, D)2m − P 2m ∈ 91,δ
2m−1
and since B is elliptic of degree 2m − 1 we obtain (2.3) and (2.4)–(2.5) follows. 2
Proof of Lemma 2.3. Let n ∈ N \ {0} and assume that for 0 < |α̃| 6 n,
|∂ α̃ x(y, η)| + hηi−1 |∂ α̃ ξ(y, η)| 6 Cα̃ hηim(α̃) , (A.2)
0
where α̃ = (α̃ 00 , α̃ 0 ) ∈ Nd × Nd and m(α̃) = −|α̃ 0 | + (1 − ρ)(|α̃| − 1).
We are going to check that the above hypothesis implies
0
|∂ α ã(y, η)| 6 Cα hηim−|α |+(1−ρ)|α| (A.3)
d0
for all α = (α 00 , α 0 ) ∈ N × Nd such that |α| 6 n.
To begin the proof we consider
0
β = (β1 , . . . , βd ) ∈ Nd , β̄ = (β̄1 , . . . , β̄d 0 ) ∈ Nd
and
σ = (αk,j )16j 6d, 06k6βj , σ̄ = (ᾱk,j )16j 6d, 06k6β̄j ,
where
00 0 0 00 0 0
αk,j = (αk,j , αk,j ) ∈ Nd × Nd , ᾱk,j = (ᾱk,j , ᾱk,j ) ∈ Nd × Nd
are such that αk,j = 0 ⇔ βj = 0, ᾱk,j = 0 ⇔ β̄j = 0. Introduce

Y Y Y Y
bσ = bk,j , b̄σ̄ = b̄k,j ,
16j 6d 06k 6βj 16j 6d 0 06k 6β̄j
316 LECH ZIELINSKI
where
bk,j = ∂ αk,j ξj if αk,j =6 0, bk,j = 1 if αk,j = 0,

b̄k,j = ∂ ᾱk,j xj if ᾱk,j =6 0, b̄k,j = 1 if ᾱk,j = 0.
Then ∂ α ã is a linear combination of terms

β
(∂ξ ∂xβ̄ a)(x(y, η), ξ(y, η))(bσ b̄σ̄ )(y, η),
where the decompositions σ , σ̄ are such that

X X X X
α= αk,j + ᾱk,j .
16j 6d 06k 6βj 16j 6d 0 06k 6β̄j
Since
β
|(∂ξ ∂xβ̄ a)(x(y, η), ξ(y, η))| 6 Cβ,β̄ hηim−|β|+(1−ρ)(|β|+|β̄|) ,
we obtain (A.3) from (A.2) noting that
−|β| + (1 − ρ)(|β| + |β̄|) +

X X X X
+ (1 + m(αk,j )) + m(ᾱk,j )
16j 6d 16k 6βj 16j 6d 0 16k 6β̄j
0
= −|α | + (1 − ρ)|α|.
Assertion (b) follows from (a) by a simple calculus. 2
Proof of Lemma 3.2. Let 0 < s 6 1/2. Then we have
|ξ 0 − ξ | < shξ i ⇒ (1 − s)hξ i 6 hξ + τ (ξ 0 − ξ )i

6 (1 + s)hξ i for τ ∈ [0; 1]
⇒ |∂ξj p0 (x, ξ 0 ) − ∂ξj p0 (x, ξ )|
6 C|ξ 0 − ξ | sup hξ + τ (ξ 0 − ξ )i−1
06τ 61
6 Cs/(1 − s) 6 2Cs. (A.4)
Now it is possible to adopt aR reasoning of [7, Chapter 18]. Instead, we may use
γs ∈ C0∞ (B(0, s)) such that γs (x) dx = 1, γs > 0, setting
Z
χj,c (ξ ) =
0
γs (hξ i−1 (ξ 0 − ξ ))hξ i−d dξ 0 , (A.5)
0j (3c/4)
similarly as in Appendix of [29]. Clearly χj,c 0

(ξ ) 6= 0 implies the existence of ξ 0 ∈
0j ( 34 c) such that |ξ 0 − ξ | < shξ i and due to (A.4), ξ ∈ 0j ( 34 c − 2Cs) ⊂ 0j (c/2)
if s is small enough to ensure 34 c − 2Cs > c/2.
Since 1 − χj,c 0
can be expressed by taking Rd \ 0j ( 34 c) instead of 0j ( 34 c) on the
right-hand side of (A.5), χj,c 0
(ξ ) 6= 1 implies the existence of ξ 0 ∈ Rd \ 0j ( 34 c)
such that |ξ 0 − ξ | < shξ i, implying ξ 0 ∈ Rd \ 0j ( 34 c + 2Cs) due to (A.4) and we
may take s small enough to ensure 34 c + 2Cs < c.
Next for s > 0 we introduce χs ∈ C0∞ (B(y0 , s)) such that χs = 1 on
B(y0 , 34 s). Then to complete the proof of (a) it suffices to take χj,c,r 0
(x, ξ ) =
R α
χ3r (x − y0 )χj,c (ξ ) and it is not difficult to show the estimates |∂ξ γs (hξ i−1 (ξ 0 −
0
ξ ))| dξ 6 Cα hξ id−|α| implying χj,c,r 0

∈ S1,0
0
.
To prove (b) we choose {0j }−d 6j 6d being a partition of Rd such that 0j0 ⊂
0
0j (2c) when 0 < |j | 6 d, 000 is bounded and set

Z
χj,c,r (x, ξ ) = χ2r (x − y0 ) γs (hξ i−1 (ξ 0 − ξ ))hξ i−d dξ 0
0j0
with s > 0 small enough. 2
Proof of Lemma 4.1. Below we have always α = (α 00 , α 0 ), ᾱ = (ᾱ 00 , ᾱ 0 ) ∈

N × Nd and we show that for (t, y, η) ∈ Cθ we have
d+1
0
|∂ α K(t, y, η)| 6 Cα hηi(1−ρ)|α|−|α | , (A.6)
where
 ∂x ∂x ∂x 
hηi
 ∂t ∂y ∂η 
K(t, y, η) =   (t, y, η).
∂ξ ∂ξ ∂ξ
hηi−1 hηi−1
∂t ∂y ∂η
Since
∂t K(t, y, η) = M(t, y, η)K(t, y, η), (A.7)
holds with a uniformly bounded matrix
 ∂(∂ p ) ∂(∂ξ p0 ) 
ξ 0
hηi
 ∂x ∂ξ 
M(t, y, η) =   (ϑ(t, y, η)),
∂(∂x p0 ) ∂(∂x p0 )
−hηi−1 − ∂ξ
∂x
it is clear that (A.6) holds when α = 0.
Next we fix n ∈ N \ {0} and assume that (A.6) holds when |α| 6 n − 1. This
assumption implies that
0
|∂ ᾱ x(t, y, η)| + hηi−1 |∂ ᾱ ξ(t, y, η)| 6 Cᾱ hηi−|ᾱ |+(1−ρ)|ᾱ|
for |ᾱ| 6 n and reasoning as in the proof of Lemma 2.3, we obtain
0
|∂ ᾱ M(t, y, η)| 6 C̃ᾱ hηi(1−ρ)|ᾱ|−|ᾱ | for |ᾱ| 6 n. (A.8)
318 LECH ZIELINSKI
Differentiating the system (A.7), we find

∂t ∂ α K(t, y, η) = M(t, y, η)∂ α K(t, y, η) + Fα (t, y, η), (A.9)
with
X
|Fα (t, y, η)| 6 2|α| |∂ ᾱ M(t, y, η)||∂ α−ᾱ K(t, y, η)|.
06=ᾱ6α
Assume now that |α| = n. Then, for 0 6= ᾱ 6 α, we can use (A.8) and (A.6) with
α − ᾱ instead of α, getting
0
|Fα (t, y, η)| 6 C̄α hηi(1−ρ)|α|−|α | , (A.10)
when α = (α 00 , α 0 ) ∈ Nd+1 × Nd is such that |α| = n. Since

∂η p0 (y, η) I 0
K(0, y, η) = ,
−hηi−1 ∂y p0 (y, η) 0 I
it is clear that (A.10) still holds if Fα (t, y, η) is replaced by K(0, y, η), hence
the Gronwall inequality ensures that (A.10) still holds if Fα (t, y, η) is replaced by
∂ α K, i.e. (A.6) still holds when |α| = n. Then Lemma 2.3 implies ∂t x = ∂ξ p0 ◦ϑ ∈
Sρ0 ((1))(Cθ ) and ∂t ξ = −∂x p0 ◦ ϑ ∈ Sρ1 ((1))(Cθ ).
(b) We write α = (α 00 , α 0 ) ∈ Nd+1 ×Nd and denote K(t, y, η) = ∂x/∂y(t, y, η).
Then (A.6) still holds and

∂ x̃
∂t ∂y x̃ ∈ Sρ (Cθ ) ⇒ |K(t, y, η) − I | = (t, y, η) 6 C|t|.
0
∂y
Therefore choosing θ > 0 small enough, we find that y → x − x̃(t, y, η) is a con-
traction with a unique fixed point y(t, x, η) for every x ∈ Rd , i.e. y → x(t, y, η) is
a bijection of Rd and x → y(t, x, η) is its inverse. We note also that |K(t, y, η)−1 −
I | 6 C|t| and ∂ α K −1 is a linear combination of terms K −n0 (∂ α1 K)K −n1 . . .
(∂ αk K)K −nk with n0 , . . . , nk ∈ N, n0 +· · ·+nk = |α|+1, α1 , . . . , αk ∈ N2d+1 \{0},
α1 + · · · + αk = α, which allows to obtain
0
|∂ α (K(t, y, η)−1 )| 6 Cα hηi(1−ρ)|α|−|α | . (A.11)
Since
∂t ỹ(t, x, η) = −(K −1 ∂t x)(t, y(t, x, η), η), (A.12)
we have |∂t ỹ| 6 C and clearly
0
|∂ α y(t, x, η)| 6 Cα hηi−|α |+(1−ρ)(|α|−1)+ (A.13)
holds when |α| = 0. Using, moreover,
∂y
(t, x, η) = K(t, y(t, x, η), η)−1 − I, (A.14)
∂x

∂y ∂x
(t, x, η) = − K −1 (t, y(t, x, η), η), (A.15)
∂η ∂η
it is clear that the estimates (A.13) hold if |α| = 1. Next we fix n ∈ Nd \ {0} and
assume that the estimates (A.13) hold when |α| 6 n. Then reasoning as in the proof
of Lemma 2.3, it is easy to estimate the derivative ∂ α of the right-hand sides of
(A.14)–(A.15) and to conclude that the estimates (A.13) hold when |α| 6 n + 1. 2
ADDITIONAL REMARKS
Until the end {qv }v∈V is a bounded subset of Sρ,δ

m
(Rd × Rd × Rd ).
(a) Let χ1 , χ2 ∈ C0∞ (Rd ) be such that s = dist(supp χ1 , supp(1 − χ2 )) > 0 and
q̃v (x, ξ, x 0 ) = χ1 (x)qv (x, ξ, x 0 )(1 − χ2 (x 0 )).
Then {Op(q̃v e−itp0 )}v∈V , Im t 60, |t |<θ(s) is a bounded subset of 9 −∞ if θ(s) > 0
is small enough. Indeed, choosing θ(s) such that θ(s)|∂ξ p0 (x, ξ )| 6 s/2 for all
(x, ξ ) ∈ Rd × Rd , we can apply the nonstationary phase estimate, writing
q̃v (x, ξ, x 0 ) = |x − x 0 − t∂ξ p0 (x, ξ )|2n qv,n (x, ξ, x 0 ),
where {qv,n }v∈V is a bounded subset of Sρ,δm
(Rd × Rd × Rd ). Then the integrations
by parts give Kt (q̃v )(x, x ) = Kt (1ξ qv )(x, x 0 ).
0 n
m0
(b) Let {av }v∈V be a bounded subset of Sρ,δ . Then we have the composition
formula
av (x, D) Op(qv e−itp0 )∗ = Op((av • qv )e−itp0 )∗ + R̃v,t , (A.16)
where {R̃v,t }v∈V , Im t 60, |t |<θ is a bounded subset of 9 −∞ , {av • qv }v∈V is a bounded
m+m0
subset of Sρ,δ (Rd × Rd × Rd ) satisfying the asymptotic expansion
X
(a • qv )(x, ξ, x 0 ) ∼ = (−i)|α| ∂ξα ā(x 0 , ξ )∂xα0 qv (x, ξ, x 0 )/α!. (A.160 )
α∈Nd
(c) Assume the existence of a constant C > 0 such that qv (x, ξ, x 0 ) = 0

for |x| + |x 0 | > C. Then {Op(qv e−itp0 )}v∈V , Im t 60, |t |<θ is a bounded subset of
B(H s+m+d+1, H s ) for every s ∈ R. Indeed, the composition formula (A.16) allows
to reduce the question to the case s = 0, m = −d − 1. However if m = −d − 1,
then Op(qv e−itp0 ) have uniformly bounded continuous kernels and form a bounded
subset of B(L2 (Rd )).
References
1. Agmon, S.: Asymptotic formulas with remainder estimates for eigenvalues of elliptic operators,
Arch. Rat. Mech. Anal. 28 (1968), 165–183.
320 LECH ZIELINSKI
2. Birman, M. and Solomyak, A: Asymptotic behaviour of spectrum of differential equations, J.

Soviet Math. 12 (1979), 247–282.
3. Boimatov, K.: Spectral asymptotics of pseudodifferential operators, Dokl. Akad. N. SSSR 311
(1990), 14–16.
4. Brüning, Z.: Zur Abschatzung der Spektral-Function elliptischer Operatoren, Math. Z. 197
(1974), 75–85.
5. Hörmander, L.: Pseudodifferential operators and hypoelliptic equations, Amer. Math. Soc.
Symp. Sing. Integrals (1967), 138–183.
6. Hörmander, L.: The spectral function of an elliptic operator, Acta Math. 121 (1968), 193–218.
7. Hörmander, L.: The Analysis of Linear Partial Differential Operators, Springer-Verlag, Berlin,
vol. 1, 2 (1983), vol. 3, 4 (1985).
8. Ichinose, W.: Propagation of singularities for a hyperbolic equation with non-regular character-
istic roots, Osaka J. Math. 17 (1980), 703–749.
9. Ivrii, V.: Precise Spectral Asymptotics for Elliptic Operators, Lecture Notes in Math. 1100,
Springer, New York, 1984.
10. Ivrii, V.: Microlocal Analysis and Precise Spectral Asymptotics, Springer-Verlag, Berlin, 1998.
11. Kumano-Go, H.: Pseudodifferential Operators, MIT Press, Cambridge, Mass., 1981.
12. Kumano-Go, H. and Nagase, M.: Pseudodifferential operators with non-regular symbols and
applications, Funkcial. Ekvac. 21 (1978), 151–192.
13. Kumano-Go, H., Taniguchi, T. and Tozaki, Y.: Multi-products of phase functions for Fourier
integral operators with an application, Comm. Partial Differential Equations 3 (1979), 349–360.
14. Levendorskii, S. Z.: Asymptotic Distribution of Eigenvalues of Differential Operators, Math.
Appl. Soviet Ser. 53, Kluwer, Dordrecht, 1990.
15. Maruo, K.: Asymptotic distribution of eigenvalues of non-symmetric operators associated with
strongly elliptic sesquilinear forms, Osaka J. Math. 9 (1972), 547–560.
16. Métivier, G.: Valeurs propres des problèmes aux limites irréguliers, Bull. Soc. Math. France
Mem. 51–52 (1977), 125–219.
17. Métivier, G.: Estimations du reste en théorie spectrale, Rend. Semi. Mat. Torino (1983), 157–
180.
18. Miyazaki, Y.: A sharp asymptotic remainder estimate for the eigenvalues of operators associ-
ated with strongly elliptic sesqulinear forms, Japan. J. Math. 15 (1989), 65–97.
19. Miyazaki, Y.: The eigenvalue distribution of elliptic operators with Hölder continuous coeffi-
cients, Osaka J. Math. 28 (1991), 935–973.
20. Miyazaki, Y.: The eigenvalue distribution of elliptic operators with Hölder continuous coeffi-
cients II, Osaka J. Math. 30 (1993), 267–302.
21. Miyazaki, Y.: Asymptotic behaviour of the spectral function of elliptic operators with Hölder
continuous coefficients, J. Math. Soc. Japan 49 (1997), 539–563.
22. Mohamed, A.: Comportement asymptotique avec estimation du reste, des valeurs propres d’une
classe d’opérateurs pseudo-différentiels sur R n , Math. Nachr. 140 (1989), 127–186.
23. Robert, D.: Sur la répartition du spectre d’opérateurs elliptiques non-auto-adjoints à coefficients
irréguliers, Osaka J. Math. 14 (1977), 593–607.
24. Taniguchi, K.: Multi-products of Fourier integral operators and the fundamental solution for a
hyperbolic system with involutive characteristics, Osaka J. Math. 21 (1984), 169–224.
25. Tsujimoto, J.: On the remainder estimates of asymptotic formulas for eigenvalues of operators
associated with strongly elliptic sesquilinear forms, J. Math. Soc. Japan 3 (1981), 557–569.
26. Zielinski, L.: Asymptotic behaviour of eigenvalues of differential operators with irregular
coefficients on a compact manifold, C.R. Acad. Sci. Paris Ser. I. Math. 310 (1990), 563–568.
27. Zielinski, L.: Valeurs propres d’opérateurs différentiels à coefficients irréguliers, Thèse de
Doctorat (1990), Université Paris 7.
28. Zielinski, L.: Asymptotic distribution of eigenvalues of some elliptic operators with simple
remainder estimates, J. Operator Theory 39 (1998), 249–282.
29. Zielinski, L.: Asymptotic distribution of eigenvalues of some elliptic operators with intermedi-
ate remainder estimates, Asymptotic Anal. 17 (1998), 93–120.
30. Zielinski, L.: Asymptotic distribution of eigenvalues for elliptic boundary value problems,
Asymptotic Anal. 16 (1998), 181–201.
31. Zielinski, L.: Spectral asymptotics for elliptic operators with Hölder continuous coefficients, 2:
The ergodic case, in: Rep. 17th Conf. Operator Theory (Timisoara 1998), to appear.
32. Zielinski, L.: Sharp spectral asymptotics and Weyl formula for elliptic operators with non-
smooth coefficients, part 2, to be published.
323
Topological Invariants of Dynamical Systems

and Spaces of Holomorphic Maps: I
MISHA GROMOV
Department of Mathematics, Institut Hautes Études Scientifiques, 35 Route de Chartres,
91440 Bures sur Yvette, France. e-mail: gromov@ihes.fr and Courant Institute, NYU, New York,
U.S.A.
(Received: 13 August 1999)

Abstract. Departing from the symbolic dynamics, we study natural group action on spaces of
holomorphic maps and complex subvarieties.
Mathematics Subject Classifications (1991): 32Hxx, 58C10.
Key words: symbolic dynamics, mean dimension, holomorphic maps, complex subvarieties.
0. Introduction
0.1. FROM X TO X0
Start from some category of spaces X and the maps between them. These can be
bare sets with no additional structure and all maps, topological spaces and con-
tinuous maps, smooth manifolds, algebraic varieties, linear or affine spaces, etc.
Then, given a group 0, we have a functionally defined 0-space X, i.e. a space with
a 0-action, namely the Cartesian power X0 thought of as the space of X-valued
functions on 0. Here the action of 0 on X0 is induced by the left action 0 on 0,
γ 0 x(γ ) = x(γ 0 γ ).
This action is called the shift and X0 is called the (full) shift space over 0 with the
alphabet X, where the basic example is 0 = Z and X consisting of finitely many
elements called letters.
0.1.1. Maps of Finite Type

There are by far more 0-maps, i.e. 0-equivariant maps X0 → Y 0 , than those
coming from maps X → Y if our category admits finite Cartesian products. In fact,
every map ϕ from the finite Cartesian power X × X × · · · × X to Y defines a 0-
| {z }
d
map X0 → Y 0 determined by the choice of a finite subset D = {δ1 , δ2 , . . . , δd } ⊂
324 MISHA GROMOV
0 as follows: each function x(γ ) goes to y(γ ), γ ∈ 0, by the usual ‘finite differ-
ence operator’ recipe,

y(γ ) = ϕ x(γ δ1 ), x(γ δ2 ), . . . , x(γ δd ) .
In other words, y(γ ) for each γ ∈ 0 is determined by the value of ϕ on the
restriction of x to the γ -translate of D,
y(γ ) = ϕ(x|γ D),
where ϕ is interpreted as a map from Xγ D = Xd to Y . In this case, we say that our
map x 7→ y is based on D and/or defined by ϕ.
Notice that this construction can be used as a definition of 0-morphisms f over
categories of certain spaces X, e.g., for algebraic varieties. But for the topological
category, there are additional continuous 0-maps X0 → Y 0 , not coming by the
way of ϕ: XD → Y . In fact, every continuous map ϕ: X0 → Y (which may
essentially depend on infinitely many x(γ ), γ ∈ 0) defines f = fϕ : X0 → Y 0 by
the same rule f : x → y for y(γ ) = ϕ(γ x).
0.2. SUBSPACES IN X0
The simplest 0-invariant subset in X0 consists of the fixed point set Fix 0 ⊂ X0
which obviously identifies with X itself, realized by the construct maps 0 → X.
More interesting subspaces in X = X0 appear as pull-backs of fixed points in Y
by 0-maps f : X → Y . One can think of such a subspace X0 = f −1 (y0 ), y0 ∈ Y ,
as the set of solutions to the ‘difference’ equation f (x) = y0 and if f = fϕ for
ϕ: XD → Y with D = {δ1 , . . . , δd } as earlier, then this equation turns into the
following system of algebraic equations denoted (ϕγ ), γ ∈ 0,

ϕ x(γ δ1 ), x(γ δ2 ), . . . , x(γ δd ) = y 0 ∈ Y . (ϕγ )
In fact, one can drop Y from this definition and start with an arbitrary subset L ⊂
XD = Xd , d = card D (corresponding to ϕ −1 (y 0 ) in the previous setting). Then
X0 = X0 (L) ⊂ X0 is defined as the space of functions x: 0 → X such that the
restriction of x to each translate γ D is contained in L, where we identify Xγ D with
XD via the correspondence γ δ ↔ δ, δ ∈ D, and where we view Xγ D as the space
of functions γ D → X. These X0 = X0 (L) ⊂ X0 are called subshifts of finite type
in X0 (where ‘finite’ refers to the finiteness of D ⊂ 0) and L is regarded as a ‘law’
distinguishing ‘legal’ function on 0.
0.2.1. Remark on Quotient Spaces

Besides taking subspaces, one may consider various 0-equivariant quotient spaces
of X0 and of the above X0 ⊂ X0 where the most attractive ones are defined by 0-
invariant equivalence relations on X0 (or on X0 ⊂ X0 ) of finite type. The simplest
TOPOLOGICAL INVARIANTS OF DYNAMICAL SYSTEMS: I 325
example is the quotient space X0 / Fix 0, where the fixed point set Fix 0 ⊂ X0 is
shrunk to a single (fixed) point. Finding more interesting 0-equivariant equivalence
relations
R ⊂ X0 × X0 ⊂ X0 × X0 = (X × X)0
of finite type is a nontrivial matter which we will not discuss at this stage.
0.3. FROM GEOMETRY TO DYNAMICS
Given our category of spaces X, take an invariant (property, theory) in this category
and try to extend it to a class of 0-spaces including X0 and subshifts of finite
type in X0 . Our extension must satisfy Inv0 X0 = Inv X and the essential (formal
functorial) properties of Inv0 must be similar to those of Inv. Besides, we want
our new invariant Inv0 to be ‘dynamical’ which expresses a vague idea of Inv0
depending on the overall behavior of the 0-orbits. For example, we wish
Inv0 (X0 / Fix 0) = Inv0 X0 ,
so that the ‘few’ fixed points of 0 should not matter.
Here is a specific example indicating what we have in mind.
0.3.1. EMBEDDING PROBLEM. Let X and Y be topological spaces where we

have a nontrivial obstruction for the existence of a topological embedding X → Y ,
e.g., S 1 6⊂ R1 or RP 2 6⊂ R3 . Does this obstruction translate to a dynamical lan-
guage and yield a nonembedding result for the 0-spaces X0 and Y 0 with their
respective product topologies and the shift actions of 0?
Of course, every 0-embedding X0 → Y 0 automatically embeds X = Fix0 ⊂
X to Y 0 = Fix0 ⊂ Y 0 and so we have a trivial ‘yes’ to our question. But if we
0
take X• = X0 / Fix0 and Y• = Y 0 / Fix0 , then the nonexistence of a 0-equivariant

embedding X• → Y• does not (seem to) immediately follow from what we know
for maps X → Y . And, truly, what we want to show is that every continuous 0-map
X0 → Y 0 identifies ‘many’ pairs of points (and thus of 0-orbits) in X0 .
0.3.2. SUBEXAMPLE. Let 0 = Z and observe that every Z-embedding between

Z-spaces, say f : X → Y , sends the periodic points of X to those of Y , i.e. the
subset Pern (X) =def Fix(nZ) ⊂ X goes to Pern (Y ) = Fix(nZ) ⊂ Y for each
n ∈ N where, obviously, all periodic points are dense in XZ . In particular, there
is no 0-embedding from X• = X0 / Fix 0 to Y• = Y 0 / Fix 0. All this is obvious
and trivially extend to all residually finite (see 1.3) groups 0 but more general 0
provide many challenging problems as we shall see later on.
0.3.3. Among 0-embeddings X → Y one distinguishes 0-homeomorphisms and,

actually, when we speak of 0-invariants, one means invariance under 0-homeo-
morphisms. Here one may have an extra structure on our spaces (e.g., a measure,
326 MISHA GROMOV
symplectic structure, complex structure, etc.) and one wishes to study (groups of)
0-homeomorphisms preserving such a structure.
0.3.4. The 0-topology can be naturally relaxed to 0-homotopy with many standard
invariants (such as homology) passing from X to X = X0 . For example, the global
homological dimension becomes Fildim(X : 0) within this framework (see 1.1.6).
0.4. MEAN ENTROPY AND MEAN DIMENSIONS
The simplest nonembedding theorem is the pigeon hole principle: there is no em-
bedding X → Y if card X > card Y for finite sets X and Y . The dynamical
version of the cardinality, or rather of the entropy =def log(cardinality), is the
mean topological entropy defined for arbitrary compact (and sometimes noncom-
pact) topological 0-spaces X (i.e. with continuous actions of groups 0) denoted
ent(X : 0) (see 1.7 for a definition). If 0 is an amenable group (see 1.3), then, not
surprisingly,
ent(X0 : 0) = ent X0 / Fix 0 = ent X (ent =)
for all finite sets X. This is a common knowledge. (Probably, something like this
must be true for sets X of infinite cardinality where the interesting 0’s are those
with card 0 > card X.) Also, one knows that
ent(X : 0) 6 ent(Y : 0) (ent 6)
if X admits a topological 0-embedding to Y or, more generally, if there is a finite-

to-one 0-map f : X → Y , (i.e. card f −1 (y) < ∞ for all y ∈ Y ). It follows
that there is no finite-to-one (not even countable to one) map f : X0 → Y 0 if
card X > card Y and the group 0 is amenable. (It is clear for all 0 that there is no 0-
embedding f : X0 → Y 0 as this would embed X = Fix0 ⊂ X0 to Y = Fix0 ⊂ Y 0
but I do not see how to exclude 0-embeddings X0 / Fix 0 → Y 0 / Fix0 for general
groups 0.)
Now, let us replace the cardinality by the topological dimension of underlying
space X which we assume at the moment being a compact metric space with finite
topological dimension. One can mimic the way one goes from ent X to
ent(X : 0) and define the mean dimension dim(X : 0) in the spirit of Lebesgue
(with the Lebesgue number of an ε-covering replacing log card (covering) appear-
ing with the entropy) for all topological 0-spaces (see 1.5). Here again
dim(X0 : 0) = dim(X0 / Fix 0 : 0) = dim X (dim =)
for most reasonable (see 1.1.5) spaces X and amenable groups 0. Furthermore,
dim(X : 0) 6 dim(Y : 0) (dim 6)

if X admits a 0-embedding to Y or, more generally, a 0-map f : X → Y with

dim f −1 (y) 6 d < ∞ for all y ∈ Y . This leads to the nonexistence of such a
map f from X0 to Y 0 if dim X > dim Y and 0 is amenable. Also, one sees in
this way that if dim X > dim Y , then X0 / Fix 0 does not 0-embed to Y 0 / Fix0 for
amenable groups 0 where the case of general 0 remains unclear.
0.5. SMOOTH SUBSHIFTS OF FINITE TYPE
Let X be a smooth manifold and L ⊂ XD , D ⊂ 0, a smooth submanifold or, more

generally, a stratified subset, e.g. an analytic subvariety in XD . One thinks of such L
as the zero set of r = codim L (sufficiently generic) equations ϕj (x1 , . . . , xd ) = 0,
d = card D, j = 1, . . . , r and then X0 = X0 (L) is given by r 0-invariant systems
of equations. So the expected mean dimension of this X0 is
?
dim(X0 : 0) = dim X − codim L. (?)
PROBLEM. Find specific sufficient conditions on L which would guarantee the

above equality.
EXAMPLE. Let X be the complex projective space CP n and L ⊂ (CP n )D =

(CP n )d , where D = {δ1 , . . . , δd } ⊂ 0, be a complex algebraic subvariety. We
shall show in 2.6.6 (using positivity of the cycle represented by L) that

dim X0 (L) : 0 > dim X − codim L (>)
for all L. Then we prove that the equality holds for (suitably understood) generic L.
Remark. Evaluating dim(X0 : 0) and, in particular, verifying (?) is not a trivial

matter even for X = Rs and linear laws L ⊂ (Rs )D since the γ -translates of the
linear equations

ϕj x(δ1 ), . . . , x(δd ) = 0, j = 1, . . . , r,
may develop unexpected linear relations. These are easy to control for such groups
as 0 = Z for instance and, to some extent, for more general unique product groups
(see 2.2.2) but the general case seems rather subtle.
0.6. SPACES OF HARMONIC MAPS AND MINIMAL VARIETIES
The most interesting spaces from our point of view appear as solutions of elliptic
differential equations over manifolds V with groups 0 acting on V . A basic ex-
ample is the space of harmonic maps V → X between Riemannian manifolds V
and X, where V is noncompact, X is compact and where V comes along with an
isometry group 0, such that V / 0 is compact. For instance, one may take V = Rn ,
where 0 is either taken to be all Rn or some lattice 3 ⊂ Rn .
328 MISHA GROMOV
The full space of the harmonic map V → X is too big and usually has infi-
nite mean dimension but it has interesting 0-invariant subspaces where the mean
dimension is finite and, sometimes, different from zero. A particular space of this
kind, denoted Xc , is distinguished by the pointwise bound on the differential of
such maps x: V → X, namely
kDxk 6 c < ∞. (∗)c
0.6.1. Upper Bound on the Mean Dimension of Xc (see 3.4)

If 0 is amenable then
dim(Xc : 0) < ∞ (?)
for all c > 0. Furthermore, if c is sufficiently large, c > c0 (V , X), then
dim(Xc : 0) 6 bcn (?)∞
for n = dim V and some constant b = b(V , X, 0), where b = a(V , X) vol(V / 0)
for discrete groups 0. Moreover,
dim(Xc : 0) → 0 for c → 0. (?)◦
Remark. If V = Rn , then (?)∞ holds true for all c > 0 as follows by an obvious
scaling argument. Probably, this remains valid for nonflat metrics on Rn invariant
under Zn but, in general, the asymptotics of dim(Xc : 0) for c → 0 should depend
on the growth rate of the group 0.
0.6.2. Nonvanishing of dim(Xc : 0) and Instantons

We shall prove in 3.6 the following:
THEOREM. Let V be a complex manifold, where an amenable group 0 acts

discretely by complex analytic transformations, such that the quotient V / 0 is a
projective algebraic variety. Then the space Xc of complex analytic maps x: V →
CP N with kDxk 6 c satisfies for all N > dimC V , and all c > c0 = c0 (V , X) > 0,
dim(Xc : 0) > b0 cn , (??)
for n = dimR V and some positive constant b0 = b0 (V , X, 0), which is of the form
a 0 (V , X) vol(V / 0) for discrete 0.
Remarks. (a) If V is Kähler, then holomorphic maps are harmonic and so

dimc (X : 0) is also bounded from above according to (?)∞ and (?)0 . In fact, these
bounds remain valid without V being Káhler as we shall see in 3.4.
(b) It seems that the strict inequality dim(Xc : 0) > 0 manifesting the abun-
dance of our maps V → X is intimately linked to the bubbling phenomenon, i.e.
the presence of instantons, highly localized solutions of our elliptic equations. Here
is a specific conjecture:
0.6.3. CONJECTURE. Let X be a complex projective manifold and look at the

space Xc of holomorphic maps x: C → X with derivatives bounded by some
c > 0. Then dim(Xc : C) > 0, if and only if X contains a rational curve.
Here (for holomorphic maps of C) the ‘if’ part of the conjecture follows from
above and also can be derived by a simple interpolation argument. On the other
hand, the ‘only if’ claim (which parallels Lang’s conjecture on hyperbolicity of X)
requires a study of ‘normal’ deformations of holomorphic curves in X which we
postpone till the second part of this paper. (At the moment, I worked out the proof
only under rather unpleasant technical assumptions.)
Remark on continuity of dim(Xc : 0). It is easy to see in many cases that the
mean dimension dim(Xc : 0) is continuous in c ∈ R+ and whenever it is positive, it
is also nonconstant as a function of c. Thus we get 0-spaces with mean dimension
taking continuous spectra of values. To see it clearer, take the case of meromorphic
functions, i.e. holomorphic maps x: C → P 1 where we bound the (spherical)
derivative by one, i.e. take X = X1 = {x | kDxk 6 1}. Then consider a lattice
3 = λZ2 ⊂ C, for λ ∈ C× , and observe (this is nearly obvious) that
dim(X : 3) = |λ|2 dim(X : C), (+)
as |λ|2 equals the volume (area in this case) of the fundamental domain of 3 in C.
Thus, by varying 3 with λ we get a continuum of mean dimensions of 3-spaces.
Next we observe that the restriction map ρλ : X → (P 1 )3 , where we evaluate
our maps x: C → P 1 at the points z ∈ 3, is injective for all sufficiently small λ.
In fact this follows from the Cauchy inequality and yields the finiteness property
(?) for the present case as

dim(X1 : C) = |λ|−2 dim(X : 3) 6 |λ|−2 dim (P 1 )3 : 3 = 2|λ|−2
(see 3.4). Now, our space X is embedded into the shift space (P 1 )3 = (P 1 )Z ,
2
where 3 = λZ2 and λ ∈ C× is small, with a continuously varying mean dimension

of the image Xλ = ρλ (X1 ) ⊂ (P 1 )3 = (P 1 )Z . Actually, dim(Xλ : Z2 ) varies in
2
the interval (0, 2], since for large λ, where the lattice 3 = λZ2 is sufficiently
rare, the restriction map ρλ : X1 → 3 becomes onto as every map 3 → P 1 can
be extended (interpolated) to a holomorphic map X: C → P 1 with kDxk 6 1
(see 3.6, where such an interpolation is used to show that dim(X1 : C) > 0).
Remark on the bound kDxk 6 1. This may look quite restrictive but, in fact,
harmonic (holomorphic) maps x with kDxk 6 1 often give a fair representation
330 MISHA GROMOV
of all harmonic (holomorphic) maps. For example, if we deal with holomorphic

(or pseudoholomorphic) maps x of C, then the AffC-orbit of every x0 : C → X
for compact X contains, in its closure, a nonconstant holomorphic map x with
kDx0 k 6 1, where AffC, where the group of transformations z 7→ λz + µ of C
naturally acts on the spaces of holomorphic maps of C. This simple remarkable
dynamical property of spaces of holomorphic maps, called Bloch–Brody principle,
will be expanded further in the second part of this paper.
0.6.4. About Residual Dimension

Let 0i ⊂ 0, i = 1, . . ., be a decreasing
T sequence of subgroups of finite index →
∞, where we emphasize the case i 0i = {id} (which makes 0 residually finite).
Then we consider subspaces Xi ⊂ X of 0i -invariant (holomorphic, harmonic etc.)
maps V → X which correspond to maps from Vi = V / 0i to X. In our case (when
we deal with harmonic maps, holomorphic maps, etc.), the ordinary dimensions
of these Xi are finite and, moreover, are bounded by const card(0/ 0i ) (see 3.4.3),
but it is unclear when the limit limi→∞ dim Xi / card(0/ 0i ) exists. If it does, it can
be called the residual dimension resdim(X : 0) and it is tempting to conjecture it
equals the mean dimension dim(X : 0) in many interesting cases.
EXAMPLE. Let V = Cn , 0i = iZ2n , i = 1, 2, . . ., and X be a projective algebraic

variety, e.g. X = CP N , N > n. If X = Xc consists of holomorphic maps x: Cn →
X with kDxk 6 c, then Xi = Xc,i are made of such maps xi from the tori Cn /iZ2n
to X. The bound kDxk 6 c obviously implies that the volumes of the images of
these maps counted with multiplicities (as well as the volumes of their graphs in
Cn /iZ2n × X) are bounded by d = const(X)(ci)2n . With this in mind, we define
the space And of ‘Abelian subvarieties in X of degree 6 d’, i.e. of pairs (A, x) where
A is an n-dimensional Abelian variety and x: A → X is a holomorphic map with
n-dimensional image whose volume counted with multiplicity is bounded by d. It
is rather obvious (see 3.4.3) that dim And 6 d const(X) and probably it is not hard
to prove the existence of the limit an = an (X) = limi→∞ d −1 dim And . Then we
define the corresponding space Yd of holomorphic maps x: Cn → X by requiring
that their graphs Gx : Cn → Cn × X have
vol Gx (B) 6 d vol B
for all unit balls B ⊂ Cn . (Actually, it would be more logical to require Vol x(B) 6
d but then one must be more careful in compactifying the resulting space of maps.)
The space Yd admits a natural Cn -invariant compactification, say Y d with the mean
dimension bounded by d const(X). (This bound follows from the first main the-
orem of the Nevanlinna theory as was pointed out to me by Alex Eremenko.) It
is not hard to show that the limit limd→∞ d −1 dim(Yd : Cn ) exists but it appears
more difficult to show this limit equals the above number an (X). Observe that a
rough bound on dim(Yd : Cn ) in terms of an for n = 1 would solve conjecture
0.6.2. On the other hand, 0.6.2 is vacuous for such spaces as X = CP N , for
instance, but the equality between the two dimensions, one referring to all maps
C → CP N and the other to iZ2 -invariant maps, does not seem obvious even for
N = 1. (Actually,√the easiest case concerns not maps of C but rather of C/Z versus
maps of C/Z ⊕ i −1Z, i = 1, 2, . . . , to CP 1 .)
0.6.5. Spaces of Subvarieties

Take a Riemannian manifold W and consider the space X of all closed subsets
M ⊂ W with the Hausdorff convergence topology on compact parts of W . Clearly,
X is compact. Notice that each isometry group 0 of W continuously acts on X,
where, obviously, dim(X : 0) = ∞ unless X/ 0 is finite.
The subsets in W worth looking at are those coming from some class of n-
dimensional subvarieties M ⊂ W which satisfy an elliptic equation (e.g. being
minimal, complex analytic, etc.) and, furthermore, are locally bounded in a suitable
sense. Then the space M of such M’s is expected to have dim(M : 0) < ∞, for
a cocompact amenable isometry group of W and this dimension should be positive
in significantly many examples. Here is a specific
THEOREM. Let W be a Hermitian manifold isometrically acted upon by a co-

compact amenable group 0. Denote by M ed the space of n-dimensional complex
subvarieties M ⊂ W , such that the intersection of M with every unit ball B in W
satisfies
Vol2n (M ∩ B) 6 d
for a given d > 0. Then

ed : 0) 6 const < ∞
dim(M
for some const = const(W, 0, d). Furthermore, if 0 is discrete and the quotient
space W/ 0 is projective algebraic, then, for 0 6 n = dimC M < dimC W , one
has
ed : 0) > const0 d n+1 ,
dim(M
for all sufficiently large d > d0 (W ) and some positive constant const0 =
const0 (W, 0) > 0.
EXAMPLE. The above applies to complex subvarieties M ⊂ CN with 0 = Z2N

and implies, for instance, that there is no Z2N -equivariant topological embedding
from Md to Md 0 if d is much (?) larger than d 0 .
Remark. This example should be taken with a pinch of salt as our proof of the
lower bound on dim(M ed : 0) is based on a 0-embedding of M ed to ([0, 1]N1 )0
while the lower bound exploits an embedding ([0, 1] ) → M
N2 0 ed .
332 MISHA GROMOV
0.6.6. Subvarieties in Compact Spaces and Residual Dimension

Along with the mean dimension, one considers the residual dimension of X refer-
ing, for example, to subvarieties in the tori Rn /i3 for a lattice 3 ⊂ Rn and i → ∞
(see 4.2).
0.7. ABOUT THIS PAPER
The present notion of mean dimension(s) arose from my attempts to geometrize

the algebraic and model theoretic conception of dimension over difference fields.
It was gratifying to see that the mean dimension distinguishes certain spaces of
holomorphic maps, thus rekindling my hopes of setting some branches of the
Nevanlinna theory into a dynamical casting. I could not trace this definition in the
literature and, apparently, this did not come up in the dynamical systems, as was
confirmed to me by Benjy Weiss with whom I was fortunate enough to discuss the
subject matter. Benjy encouraged me by showing his interest in the mean dimen-
sion (actually, it was Benjy who suggested the ‘mean dimension’ terminology) and
he immediately generated a flow of dynamical ideas, including several conjectures
relating the mean dimension and entropy. Many of his conjectures have already
turned into theorems which appear along with many other results in [Lin-Wei]
and [Lin]. Then I had an opportunity to discuss the holomorphic part of this paper
with Mario Bonk and Alex Eremento. Alex explained to me several essential points
on normal spaces and professionally sharpened the inequalities on the dimension of
the spaces of meromorphic maps (see his survey paper [Ere]). More recently, I had
a pleasure of talking to Michael McQuillan about the problems related to Lang’s
conjecture which made me more confident in my mean-dimensional version of it.
Part I of our paper focuses on elementary properties of the mean dimension and
on illustrative examples. More technical discussion is postponed until Part II.
1. Mean Dimension in Various Categories of 0-Spaces

1.1. WIDTH AND DIMENSION
A map f : X → P , where X is a metric space, is called an ε-embedding if f does

not identify points in X with distances > ε. In other words,
Diam f −1 (p) 6 ε for all p ∈ P .
Then, following Uryson, we define Widimε X as the minimal number k, such that
X admits a continuous ε-embedding to a k-dimensional polyhedron P . Clearly,
Widimε is monotone decreasing in ε.
1.1.1. The basic example of evaluation of this ε-dimension is the following:
LEBESGUE LEMMA. The unit cube [0, 1]N ⊂ RN has

Widimε [0, 1]N = N for all ε < 1.
Consequently,
Widimε RN = N for all ε > 0.
Here is a more general (and slightly less precise) Widim inequality:
1.1.2. Widim Inequality

Let B be the unit ball in an N-dimensional Banach space. Then
Widimε B = N for all ε < 1. (∗)
Proof. The inequality Widimε B 6 N − 1 trivially implies that FilRad(∂B) 6

ε/2 (compare App. 1 in [GroFRM ]). On the other hand, the boundary sphere S N−1 =
∂B with the induced metric has FilRad S N−1 > 1/2 by the argument in 1.2.3
of [GroFRM ] since every k-tuple of points in this S N−1 with mutual distances < 1
canonically (and obviously) spans a (k − 1)-simplex in S N−1 . 2
Remark. The above will be used in 2.4 for evaluating the mean dimension of
(sub)-linear subshifts Y ⊂ B 0 ⊂ (Rs )0 , where we shall need another lemma:
1.1.3. TRIVIAL LEMMA. Let Y be a closed subset in a Banach space X and let
p: X → RN be a bounded linear operator. Then, for arbitrary metrics on Y and on
p(Y ) ⊂ RN compatible with their topologies, one has Widimε Y > Widimε p(Y )
for all ε > 0.
Proof. As the fibers of the map p: Y → p(Y ) are all nonempty convex, there is
a continuous section, i.e. a map q: p(Y ) → Y such that p ◦ q = Id: Y → Y . Thus
one has Widimε Y > Widimε qp(Y ) > Widimε p(Y ). 2
1.1.4. OPEN QUESTIONS. The Widim inequality allows a lower bound on Widimε
of the intersection of a linear subspace Y in a Banach space X with the unit ball,
Widimε Y ∩ B > dim Y for ε < 1 (∗)
(compare 2.6). Then we wish to have a similar inequality for nonlinear subvarieties
Y ⊂ X. For example:
Does (∗) hold true for X = CN and Y being a complex analytic subvariety
passing through the origin?
We would not mind (∗) with a slightly smaller ε > 0 but the answer is not
even known for ε = εN > 0. On the other hand, it is not hard to prove (∗) with
ε depending on the degree of Y in the case Y is complex algebraic. In fact, (∗)
holds true with ε = ε(Vold (Y ∩ 2B)), d = dimR Y for all minimal subvarieties
in R2N by the usual compactness argument. It would be interesting to make such
an argument work uniformly for all dimensions and thus applicable for evaluating
of the mean dimension of (local) algebraic subvarieties in (C1 )0 (compare 2.5).
334 MISHA GROMOV
On the other hand, one may ask on the possible range of Widimε for given ε on a
given class of subvarieties and then one is tempted to extend this question to other
‘slicing invariants’ of Y ∩ B defined in App. 1 of [GroFRM ].
It seems Widimε has not been evaluated even for simple convex subsets in Rn ,
e.g. for the simplex 1n−1 = {xi > 0, 6xi = 1}, where one expects (maybe too
navely) that
Widimε 1n ∼ constε n.
P
Another interesting example is the Euclidean ball B`2 = { ni=1 xi2 6 1} whose
Widimε is to be measured with respect to the sup-product metric (with the corre-
sponding norm kxk`∞ = supi=1,...,n |Xi |). More generally, one asks what is
Widimε B`p with respect to the `q -norm in Rn ?
1.1.5. It is clear that Widimε X < ∞ for all compact metric spaces X and all ε > 0
but it may become infinite for noncompact spaces X (where, in fact, the definition
must be modified by replacing Diam f −1 (p) by lim supU →p Diam f −1 (U ) where
U runs over the neighbourhoods of U in P ) and this inequality is strict. It is also
clear that Cartesian product X1 × X2 with the sup-product metric, that is

dist (x1 , x2 ), (x10 , x20 ) = max dist(x1 , x10 ), dist(x2 , x20 ) ,
satisfies the product inequality

Widimε (X1 × X2 ) 6 Widimε X1 + Widimε X2 .
It follows, that Widimε is also subadditive for taking maxima of metrics on the
same space X,
Widimε (X, dist) 6 Widimε (X, dist1 ) + Widimε (X, dist2 )

for dist = max(dist1 , dist2 ).
Warning. One should be careful with the additivity of Widimε for Cartesian
products. In fact, even the Lebesgue dimension is not always additive, but the extent
of the nonadditivity is completely clarified by the work of Dranishnikov (see [Dra])
who kindly explained this to me.
1.1.6. Remarks on covε and Fildimε . The ε-dimension Widimε X, as a function

of ε carries the same information about the geometry of X as the totality of its
Uryson’s widths (see [GroNLS ]). A more traditional and essentially equivalent defi-
nition of ε-dimension is the Lebesgue covering number Lebε X, that is the minimal
intersection multiplicity of the ε-covers of X minus one. We prefer Widimε as this
leads to interesting variations of the theme in the spirit of metric geometry such as
the global ε-dimension Fildimε X. The latter is defined as the maximal dimension
of cycles C ⊂ X with FilRad C > ε, i.e. nonbounding in any metric extension

Y ⊃ X with supy∈Y dist(y, X) 6 ε (compare[GroFRM ]).
1.2. DEFINITION OF Widimε (X : {i }) AND Widimε (X : 0)
Let X be a metric space and a group 0 act on X. We assume 0 is given a proper

left invariant metric, where ‘proper’ means that the balls B(γ , R) = {γ 0 ∈ 0 |
dist(γ 0 , γ ) 6 R} are compact for all R < ∞. Also, we fix a left invariant (Haar)
measure on 0, denote || = measure () and observe that || < ∞ for all
bounded (with respect to the metric) domains in 0.
Our basic examples are Lie groups, such as 0 = Rn with the usual metric
and measure, as well as discrete finitely generated groups 0 with given generators,
γj , . . . , γi , . . . , γk where the word metric is defined by setting dist(id, γ ) equal to
the length of the shortest words in γi representing γ and where || = card .
We denote by |x − x 0 |γ the γ -translate of the original metric on X, denoted
|x − x 0 |, and assume that the identity map (X, |x − x 0 |γ1 ) → (X, |x − x 0 |γ2 ) is
uniformly continuous for all γ1 , γ2 ∈ 0 where the implied continuity modulus
depends only on dist(γ1 , γ2 ). In other words, the action of 0 is assumed uniformly
continuous on X. We define the metrics |x − x 0 | on X for all bounded ⊂ X
as |x − x 0 | = supγ ∈ |x − x 0 |γ and let X = (X, |x − x 0 | ). Then we look at
Widimε X as a function on bounded subsets ⊂ 0 and observe that this function
is subadditive according to the inequalities in 1.1.5. This implies, for amenable (see
below) sequences i ⊂ 0, that the limit
Widimε (X : {i }) = lim |i |−1 Widimε i
exists and does not depend on a sequence i (see 1.3.5), exactly as it happens to the
entropy (see [Orn-Weis]). Then we use this limit for the definition of Widimε (X :
0) (see 1.4).
1.3. AMENABILITY
Given a subset ⊂ 0 we define its ρ-boundary ∂ρ ⊂ 0 for all ρ > 0 as

the set of those γ ∈ 0 for which the ball B(γ , ρ) intersects as well as the
complement 0\. Then a sequence i ⊂ 0 is called amenable (or Fölner), if
|∂ρ i |/|i | → 0 for i → ∞ and each ρ > 0. In other words, the ρ-boundary of
i is ‘asymptotically negligible’. Notice that, on the one hand, this definition uses
no group structure but rather the metric and the measure on 0. On the other hand,
the amenability of a sequence does not depend on the choice of a Haar measure
and of (proper left invariant) metric on 0.
A group 0 is called amenable if it admits an amenable sequence i ⊂ 0. (If 0
is discrete or, more generally, unimodular, this equivalent to the classical definition
of amenability where every continuous action of 0 on a compact space is required
336 MISHA GROMOV
to have an invariant measure. Actually all amenable groups we consider in this

paper may be assumed to be unimodular and so one should not be bothered by the
discrepancy between the two definitions.)
1.3.1. Ornstein–Weiss Lemma (see [Orn-Weis])

Let h() be a positive function defined on bounded subsets ⊂ 0 such that
(a) h is subadditive, i.e.
h(1 ∪ 2 ) 6 h(1 ) + h(2 ) (∗)
for all pairs of bounded subsets 1 and 2 in 0.
(b) h is invariant under 0,
h(γ ) = h(), for all γ ∈ 0.
Then the limit
lim h(i )/|i | (?)
i→∞
exists for every amenable sequence i ⊂ 0.
Remark. (a) Clearly, the existence of the limit for all amenable sequences
implies its independence of a choice of a sequence.
(b) if h() is monotone increasing for 0 ⊃ , then it suffices to assume (∗)
only for disjoint subsets 1 and 2 .
Sketch of the Proof. Take two subsets 0 and in 0, where will be eventually
taken much larger than 0 , and consider some translates γi 0 ⊂ 0, i = 1, 2, . . . ,
such that:
(a) all γi 0 are contained in ; S
(b) the intersection of γi (0 ) with the union U0i−1 = i−1
j =1 γj 0 satisfies

(γi 0 ) ∩ U i−1 6 ε|0 | (∗)ε
0
for a given ε > 0.

We take a maximal sequence of translates γi 0 , i = 1, . . . , k, satisfying the ε-
packing conditions (∗)ε for all i and estimate from below the measure of the result-
ing union U0k ⊂ as follows. Denote by ρ0 the diameter of 0 , i.e. sup dist(δ, δ 0 )
for δ, δ 0 ∈ 0 and let α denote the relative amenability constant, i.e. α0 = α(, 0 )
= |∂ρ0 |/||. We claim that
|U0k |/|| > ε(1 − 2α0 ). (+)ε
To see this, let + ⊂ 0 consist of those γ for which the intersection γ 0 ∩ is

nonempty and − ⊂ 0 consist of γ , where γ 0 ⊂ . It is convenient to assume
at this point that id ∈ 0 . Then + is contained in the ρ0 -neighbourhood of , i.e.
in ∪ ∂ρ0 , while − contains the ρ0 -interior of , i.e. the complement \∂ρ0 .

Thus |− |/|+ | > 1 − 2α0 .
On the other hand, obviously,
Z
|U0k ∩ γ 0 | dγ = |U0k | |0 |
+
and so
Z
− −1
| | |U0k ∩ γ 0 | dγ 6 |U0k | |0 | |+ |−1 (1 − 2α)−1
−
6 |U0k | |0 | ||−1 (1 − 2α0 )−1 . (1)
Next, by the maximality of k, (∗)ε must be violated for all γ ∈ − , i.e.
|U0k ∩ γ 0| > ε|0 |
for all γ ∈ − and thus
Z
−1
|− | |U0k ∩ γ 0 | dγ > ε|0 |.
−
Hence,
ε 6 |U0k | ||−1 (1 − 2α|−1
and (+)ε is proven.
Now we are ready to prove the existence of the limit (?) by adopting the clas-
sical (and trivial) argument establishing convergence of h(t)/t for sublinear func-
tions h(t). Denote by `− the lower limit
lim inf h(i )/|i |,
i→∞
and take some i1 , i2 , . . . , is among i such that

(a) the ratios h(i )/i are all close to `− , say
h(i )/|i | 6 `− + ε
for a given ε > 0;
(b) the relative amenability constants α(iµ , iν ) are very small compared to ε s
for all iµ < iν ;
(c) the number s is very large.
Then we bound the ratio h()/ for all sufficiently large where the relative
amenability constants α(, iµ ) are small. To do this we start with the above ‘ε-
packing’ of by is (playing the role of 0 ). The remaining part 0 = \ ∪ γi is
has measure ≈ (1 − ε) and its ρ-boundary equals the union of these of and the
translates γi is . Thus, the relative amenability constants α(0 , iµ ) remain small
for µ < s and we can ‘ε-pack’ 0 by translates of is−1 . We keep doing this and
338 MISHA GROMOV
finally cover much of by translates of iµ , namely the union of all these translates
has total measure at least (1−(1−ε)s )(1−2α)s ||, where α is the upper bound on
the relative amenability constants. Since (we may assume) α is much smaller than
ε s , we cover almost all of . On the other hand, our covering is (1 − ε)-efficient,
i.e. the total measure of our translates does not exceed (1 + ε) according to (∗)ε .
Thus, the union of all our translates say U ⊂ has h(U ) bounded by something
of the order `+ + 2ε. On the other hand, the complement \U has small measure
and retains some ‘amenability’ having |(\U ) ∪ ∂1 (\U )| also small, say 6 ε. It
follows, by subadditivity of h, that h(\U ) is bounded by something of the order
of ε|| and h() is bounded by `−|| + O(ε)||. This yields the Ornstein–Weiss
lemma.
1.3.2. Euclidean Example

Let 0 = Rn and i be Euclidean i-balls for i = 1, 2, . . . . Then the above some-
what simplifies as large balls can be efficiently packed by smaller ones without any
overlaps at all. (This is especially useful when we deal with superadditive functions
such as maximal degrees of 1-Lipschitz maps i → S n , see [G-L-P], § 2.)
1.4. EXISTENCE OF Widimε (X : 0) FOR AMENABLE AND NONAMENABLE 0
We continue 1.1.6 and 1.3.1 and define
Widimε (X : 0) = lim Widimε (X : {i })

i→∞
with any amenable sequence i ⊂ 0.

In general, if we do not assume amenability, we set
Widimε (X : {i }) = lim inf |c |−1 Widimε Xi

def
i→∞
for all sequences i ⊂ 0 with µ(i ) → ∞. And if we want to eliminate i , we

consider all exhaustions {i } of 0 and take the infimum of Widimε (X : {i }) over
all exhaustions. This can be regarded as Widimε (X : 0) which is equal to the above
if 0 is amenable, as a simple reasoning shows. But we are not seriously concerned
with keeping our definition independent of i as all our considerations are as good
for one sequence of ’s as for another.
1.5. LETTING ε → 0 AND DEFINING dim(X : 0)
The above mean ε-dimensions Widimε (X : {i }) and Widimε (X : 0) are monotone
decreasing in ε. Thus, we can go to the limit and set
dim(X : {i }) = Widim(X : {i }) = lim Widimε (X : {i })

ε→0
and
dim(X : 0) = Widim(X : 0) = lim Widimε (X : 0)

ε→0
if we want to be i -free.
Also, we observe that this definition makes sense for every (not necessarily
invariant) subset Y ⊂ X (as we may work with the metrics |x − x 0 | restricted to
Y ) and we shall be using this for compact subsets Y ⊂ X.
If X is itself a compact metric space, then the above definition of Widim does
not depend on the original metric |x − x 0 | in X. In general, one could make things
invariant by first taking supY Widim(Y : {i }) over all compact Y ⊂ X and then
taking infimum over all metrics |x − x 0 | on X compatible with the topology of X
and such that the action of 0 on X is uniformly continuous. (We shall return to this
later on when it becomes relevant.)
1.5.1. Topological Invariance of Mean Dimension

If X is a compact space then, clearly, the mean dimension Widim(X : {i }) does
not depend on the choice of the original metric |x − x 0 | in X. In fact, continuity of
the identity map (X, |x − x 0 |old ) → (X, |x − x 0 |new ) implies uniform continuity for
the metrics |x −x 0 |old 0 new
γ and |x −x |γ simultaneously for all γ ∈ 0 and consequently
for |x − x 0 |old 0 new
and |x − x | . This gives a bound on Widimε
new
in terms of Widimold
δ
for some δ = δ(ε) and as ε → 0 we arrive at the equality Widimnew = Widimold in
the limit, since δ(ε) → ε for ε → 0.
1.5.2. Monotonicity of Widim

Clearly every 0-invariant subspace Y ⊂ X has Widim(Y : {i }) 6 Widim(X :
{i }). In fact, as we mentioned earlier, Widim(Y : {i }) makes sense for arbitrary,
not necessarily invariant, subsets Y ⊂ X as all we need are our metrics |y − y 0 |γ
on Y and these come by just restricting the metrics |x − x 0 |γ from X to Y ⊂ X for
all γ ∈ 0. Then obviously,
Widim(Y1 : {i }) 6 Widim(Y2 : {i })
for all Y1 ⊂ Y2 ⊂ X and all sequences i ⊂ 0. In particular,
dim(Y : 0) 6 dim(X : 0)
if Y admits a 0-equivariant embedding to 0.
1.6. ON ISOMETRIC ACTIONS ON BANACH SPACES
There are certain topological spaces X, which admit weak compactification, i.e. a
compact topological space X• along with a bijective continuous map e: X → X• .
340 MISHA GROMOV
For example, the unit ball {kXk 6 1} in each Banach space is like that. Clearly, if
X• exists it is unique up to homeomorphism.
Now, let X come along with an action of 0 and let weak compactification refer
to a compact 0-space X• with a bijective continuous 0-equivariant map X → X• .
This (X• , 0) is also (obviously) unique, if it exists, and its 0-invariants, such as
dim(X• : 0) can be regarded as invariants of (X, 0).
BASIC EXAMPLE. Let 0 isometrically act on a Banach space and thus on the unit
ball X in this space. A 0-invariant weak compactification is obvious for reflexible
spaces and it also exists for some (all?) other examples, such as `∞ (0). Then one
may speak of
def
dim(X : 0) = dim(X• : 0).
It is clear, that
dim(X : 0) = s
for X being the unit ball in the `∞ -space of bounded functions 0 → Rs (0 is
discrete here) and that this dimension 6 n for all other `p -spaces. But I could
not decide if it is actually positive for p < ∞ (where the problem is related to
evaluation of `∞ -width of `p -balls, compare 1.1.4) and nontrivially depends on p.
(If so, this would imply the spaces (X, 0) are mutually 0-nonhomeomorphic for
different p, which, I guess, is unknown for infinite groups 0.) This problem, on the
one hand, and the idea of the Von Neumann dimension, on the other hand, lead to
the following modification of our dim(X : 0) (see 1.12 – 1.12 3).
1.6.1. Definition of dim(X : 0)`p

Let us replace the sup-product distance |x − x 0 | from . . . by the `p -distance,
Z 1/p
0 0 p
|x − x |,`p = |x − x |γ dγ

and then repeat everything with |x − x 0 |,`p instead of |x − x 0 | . Notice that the
resulting dimension is not a topological 0-invariant, it is only a Lipschitz invariant
(and Hölder ‘covariant’ in an obvious sense). This is not so bad if we speak of
isometric actions on (balls in) Banach spaces (where even the linear Lipschitz in-
variance is a nontrivial issue) but our definition needs an adjustment to this case. It
seems reasonable to consider all compact convex metric 0-spaces X• admitting bi-
jective (surjective?) Lipschitz linear 0-maps X → X• , and take sup dim(X• : 0)`p
over all such X• ‘under’ X. (And as the discussion became linear, one might try
more manageable linear widths instead of the topological one.)
1.7. REMARKS ABOUT ENTROPY, COVERINGS , ETC .
Our definition of the mean dimension mimics that of the topological entropy where,
instead of our Widimε X, one uses entε X = log covε X, where covε is the minimal
number of the open subsets in X of diameter 6 ε needed to cover X. In fact,
one can avoid any metric in the definition of both invariants ent(X : {i }) and
dim(X S: {i }) by a direct appeal to (sufficiently fine) finite
T open covers of X, say
X = ν Uν and the associated covers by the intersections γ ∈ γ (Uν ). This defin-
ition of the mean dimension has an advantage of being applicable to nonmetrizable
spaces and it is adopted in [Lin-Wei]. We choose here Widim as it is easier on
the level of notations and also more flexible when it comes to generalizations. For
example, our definition does not truly need any action: every family 1 of metrics
|x − x 0 |δ , δ ∈ 1, on X will do. Such a situation naturally comes up in the study of
spaces X of X-valued functions over a given background space 1 replacing 0 in
the example of X = X0 . Here each point δ ∈ 1 gives rise to a metric on functions
x(δ) via some weight function w(δ, δ1 ) on 1 × 1 by the formula
|x − x 0 |δ = sup w(δ, δ1 )|x(δ1 ) − x 0 (δ1 )|X ,
δ1 ∈1
0
where |x − x |X refers to a preassigned metric on X. Typically, 1 itself is a metric
space (e.g. a graph as in [Gro]ESAV ), and
w(δ, δ1 ) = exp −β dist1 (δ, δ1 ).
‘Microscopic’ observations. One can think of a subset ⊂ 0 (or more generally
⊂ 1) as a ‘microscope’ applied to the metric space X = (X, |x − x 0 |) and en-
larging its visual image to the greater size X = (X, |x −x 0 | ), where the resolving
power of depends on the presence of transformations γ : X → X, γ ∈ , which
expand the original metric in X. This expansion brings invisibly small geometric
details of X = (X, |x − x 0 |) to the observable scale ε where we have a vari-
ety of ‘macroscopic’ geometric techniques at our disposal (see [GroFRM , GroAI ]
and [GroPCMD ]). The magnification may be highly nonuniform in different direc-
tions and so when we eventually send ε → 0 we arrive at a new ‘non-isotropic’
image of X quite different from the original (X, |x − x 0 |) (compare § 4.10 in
[GroCC ]). Thus various ‘macroscopic’ invariants discussed in the above-cited pa-
pers (e.g. Widimε X, Fildimε X, etc.) are getting transported from the geometric
realm to the domain of topological dynamics.
1.8. MEAN MINKOWSKI DIMENSION
This dimension is defined for invariant sub-spaces Y of a topological 0-space X

with a Borel measure µ on X as follows.
Let U ⊃ Y be a (noninvariant!) neighbourhood T of Y in X and consider the
intersection of the γ -translates of U , say Ui = γ ∈i γ U . Then pass to the limit
1/|i |
MU = lim sup µ(Ui ) (∗)
i→∞
342 MISHA GROMOV
and finally let
Min dim(Y : {i }) = inf MU ,

U
where U ⊂ X runs over all neighbourhoods of Y ⊂ X.
1.8.1. Motivating Example. Let (X1 , µ0 ) be a compact probability space and

Y ⊂ X be a closed subset. Then the subset Y = Y 0 ⊂ X = X0 has Min dim(Y :
{i }) = µ0 (Y ) for amenable sequences i . This directly follows from the defini-
tions.
1.8.2. Measuring Noninvariant Subsets Y ⊂ X. Instead of translating U we

may transport a given metric |x − x 0 | on X and define Uε i as the intersection of
the ε-neighbourhoods of Y with respect to the metrics |x − x 0 |γ for γ ∈ i . Then
we take the limit Mε with Uε i substituting Ui in (∗) and finally let ε → 0. The
resulting version of the Minkowski dimension (obviously) reduces to the above
Min dim for closed invariant subsets in compact probability spaces X.
1.8.3. Variation. Rather than intersecting the ε-neighbourhoods for the metrics
|x−x 0 |γ , one could take the ε-neighbourhood with respect to the metric |x−x 0 |i =
supγ ∈i |x−x 0 |γ . This may be only smaller than Uε i and so the resulting dimension
is smaller than Min dim. (Probably, there are easy examples where it is strictly
smaller.)
1.8.4. Smooth Remark. If X is a compact smooth manifold with a 0-action then

one can apply the above to a smooth (not necessarily invariant) measure µ on X. In
particular, one may speak of Min dim({x} : {i }) for all points x ∈ X and observe
that the topological entropy is (obviously) constrained by the numbers
M+ = sup Min dim({x} : {i }) and

x∈X
M− = inf Min dim({x} : {i }) as follows,

x∈X
− log M+ 6 topent(X : {i }) 6 − log M− .
1.8.5. Minkowski Dimension and Coentropy. In many examples where µ is an

invariant measure of the maximal entropy and the topological entropy is finite,
the Minkowski dimension equals exp(topent(X : 0) − topent(X : 0)). Further-
more, there are easy examples where topent(X : 0) = ∞ but (X, Y, 0) can be
approximated by actions with bounded entropy, say (Xi , Yi , 0), such that

exp topent(Yi : 0) − topent(Xi : 0) → Min dim(Y : 0),
i→∞
where the notation Min dim(Y : 0) refers to a suitable exhaustion {i } of 0.

1.8.6. About Examples. The mean Minkowski dimension is invariant under

measure preserving continuous maps, e.g. homeomorphisms, α: X0 → X where
Y 0 = α −1 (Y ). Constructing such maps is an interesting problem which makes
sense in every geometric category of X’s, where one is especially interested in
the structure of the group Aut X0 consisting of invertible maps X0 → X0 of
finite type with inverse also being of finite type. Besides right translations by 0
and automorphism in Aut X acting on X0 in an obvious way, one has two general
possibilities.
I. Triangular maps. The simplest instance of this appears where X is split,
say X = Y × Z. Here every map ϕ: Z 0 → (Aut Y )0 of finite type defines an
automorphism of X0 by (y, z) 7→ (ϕ(z)(y), z).
II. Markers. The idea is similar to the above with Aut Y replaced by Aut Y D for a
finite subset D ⊂ 0 (or a finite collection of these). Such D, as well as its translates
in 0, are distinguished by insisting on certain values of z on these D’s. If these D’s
happen to be mutually disjoint, then suitable automorphisms of Y D parametrized
by z give us automorphisms of X0 . All this has been carefully studied for shifts S Z
and finite S (see [Hed]) and we shall return to the general case in the second part
of this paper.
III. Sometimes one can ensure invertibility of a map by an implicit function
argument but then the resulting inversion is, typically, of infinite type.
IV. If X is a smooth manifold, one may speak of 0-invariant vector fields on
X0 of finite type and study the corresponding flows (which may be only partially
defined). For instance, if X is a symplectic manifold, then every function (local
Hamiltonian) h: XD → R defines such a flow. (We shall return to this and will
study the corresponding symplectic geometry in the second part of the paper.)
1.9. PROJECTIVE AND LEGAL DIMENSION IN X0
Consider a subspace Y ⊂ X = X0 and define its dimension using natural pro-

jections X0 → X ⊂ 0, (corresponding to restriction of functions from 0 to
subsets ⊂ 0) as follows. Let Y |i denote the image of Y under our projection
X0 → Xi and set
prodim(Y : {i }) = lim inf dim(Y |i )/|i |

i→∞
for every sequence of bounded subsets i ⊂ 0 with |i | → ∞.

This projective dimension looks more approachable than dim(Y : 0) =
Widim(Y : 0) and sometimes the two dimensions are known to be equal. In any
case, we have the following:
344 MISHA GROMOV
1.9.1. PRO-MEAN INEQUALITY. If X is compact, then every closed (not

necessarily invariant) subset Y ⊂ X0 satisfies
dim(Y : {i }) 6 prodim(Y |{i })
for every amenable sequence i ⊂ 0.

Proof. The projection from Y to the Xi +ρ (where i + ρ ⊂ 0 denotes the
ρ-neighbourhood of i ) is an ε-embedding with ε = ε(ρ) → 0 for ρ → ∞ and,
clearly,
dim Y | i > (dim Y | i + ρ) − |∂ρ i | dim X. 2
1.9.2. Legal Dimension. Let Y ⊂ X0 be an invariant subset of finite type

defined by a law L ⊂ XD , D ⊂ 0. Consider all translates γ D in 0 which are
contained in a given subset and let L ⊂ X consist of L-legal functions on ,
i.e. of those x: → X where the restriction of x to each γ D ⊂ is contained in
L (where, as earlier, γ D is identified with D and Xγ D with XD ). Then define
legdim(Y : {i }; L) = lim inf dim Li /|i |,

i→∞
where i is a sequence of subsets in 0 (which is assumed amenable in most

applications).
1.9.3. On Nontopological Spaces. If 0 is a discrete group where bounded sub-

sets D are finite, then the definition of legdim makes sense in every category with
(finite!) Cartesian products and a notion of dimension (or rank). For example, this
applies to linear and affine spaces over an arbitrary field and up to a certain extent
to moduli over more general (commutative and noncommutative) rings. Also, one
may use this definition for (pro)-algebraic varieties over an arbitrary field and also
for analytic varieties over a local field.
1.9.4. On Subspaces Y ⊂ X0 of Infinite Type for Metric Spaces X. If Y is a

subshift of infinite type, the projections Y → X may be easily onto (an open
subset in X ) even for relatively small Y (e.g. for X = Rs and Y being a generic
infinite-dimensional linear subspace) and so the ordinary dimensions of the images
do not tell us much. It is more useful to take the Widimε of these images Y |
which works well for example, for the space `p (0, Rs ) and Y = B ∩ Y0 , where Y0
is a 0-invariant linear subspace in our `p and B is the unit `p -ball. Here it seems
reasonable to evaluate Widimε Y | with respect to the `p -norm on `p (, Rs )
as is suggested by the `2 -case where this leads to the Von Neumann dimension
(see 1.12).
1.9.5. On Invariance of legdim and Introduction of stablegdim. The definition

of legdim depends not only on Y = Y (L) but also on the defining law L ⊂ XD
although in most cases the dependence on L is illusory.
In fact legdim = prodim in many cases (see Section 2) and it is useful to bring
in an intermediate notion of stable legal dimension, denoted stablegdim(Y : {i }),
where i is increasing sequence of subsets in 0. To define this we project Lj to
Xi for all j > i and let Lij ⊂ Xi denote the images of these projections. Then
we set stablegdim(Y : {i }; L) = lim infi→∞ limj →∞ dim Lij /|ij |. Observe
that this stablegdim extends to nontopological categories in most cases where it
is possible for legdim and this sometimes allows such an extension for prodim (see
below).
1.9.6. Elementary Inequalities. It is clear that

prodim 6 stablegdim 6 legdim
(where, recall, meandim 6 prodim for compactT Y ).
Also observe that the intersection Li∞ =def j >i Lij equals the projection Y |
i of Y to i for compact subspaces Y ⊂ X0 . Hence,
prodim Y = stablegdim Y, (∗)
provided the dimension is stable under countable intersections of subsets in our
category. This is so, for instance, for compact complex analytic varieties by the
Noether intersection property: every decreasing family of compact complex spaces
stabilizes.
Thus we have the following simple
PROPOSITION. Let X be a complex analytic variety and L ⊂ XD be a compact

subvariety. Then Y = Y (L) ⊂ X0 satisfies the above equality (∗).
Remark. The point of this is our evaluation of some dimension of a ‘trans-

cendental’ object, our Y , in terms of ‘elementary’ ones, i.e. Lij .
On Extension of Prodim to Nontopological Categories. The equality Y | =

Li∞ remains valid in many algebraic categories, (e.g. for complex algebraic va-
rieties and saturated models of first order theories in general) and if we have a
notion of dimension in our category which passes to countable intersections of
varieties (as it happens, for instance, to countable intersections of constructible
subsets in K N for an uncountable algebraically closed field K), then we can define
prodim Y (L) for laws L ⊂ XD in our category.
1.9.7. On Stable Laws. A law L ⊂ XD is called stable if there exists ρ0 , such

that the image of the projection from L+ρ to L does not depend on ρ for all ρ >
ρ0 and all bounded ⊂ 0, where, recall, +ρ ⊂ 0 denotes the ρ-neighbourhood
of in 0. Clearly, if L is stable, then
prodim(Y : {i }) = legdim(Y : {i }; L) = stablegdim(Y : {i }; L)
for all amenable sequences i ⊂ 0.
346 MISHA GROMOV
PROBLEM. Find less restrictive conditions ensuring the above equalities between
different dimensions. (See Section 2 for practical results in this direction.)
1.10. RESIDUAL DIMENSION
Given a discrete subgroup 00 ⊂ 0, we consider the fixed point set Fix 00 ⊂ X in a

given 0-space X. For example, if X = X0 , this Fix 00 consists of all 00 -invariant
functions 0 → X which can be identified with functions 0/ 00 → X.
We are especially interested in the case where 00 is of finite covolume, i.e. when
the Haar measure |0/ 00 | is finite. In this case we may expect dim Fix 00 < ∞ and
so we set
resdim X/ 0i = lim inf(dim Fix 0i )/|0/ 0i |
i→∞
for every sequence of discrete subgroups 0i ⊂ 0 of finite covolumes with |0/ 0i |

→ ∞.
The most interesting case is where the spaces 0/ 0i converge to 0, i.e. if for
each bounded subset ⊂ 0 the intersection ∩ 0i consists of {id} for all
i > i0 = i0 (). Recall that a discrete group 0 admitting such a sequence of
0i is called residually finite, and many residual finite groups are far from being
amenable. Such are the free groups and most finitely generated subgroups in the
linear group GLn R.
What may limit the applicability of the residual dimension is absence of a
sufficient amount of periodic (i.e. 0i -fixed) points. However, if L ⊂ XD is a
strongly stable law (see 7.E.000 in [GroESAV ] and below) then periodic points are
dense in Y (L) ⊂ X for residually finite groups 0 and resdim = prodim if 0 is also
amenable. This follows by the argument in 7.E.00 in [GroESAV ].
Definition of Strong Stability. Call L strongly stable if there exists ρ0 > 0, such
that the following condition (locρ0 ) is sufficient for extendability of a function
X0 : 0 → X to our x: 0 → X belonging to Y (L) ⊂ X0 ,
(locρ0 ) For every ρ0 -ball B ⊂ 0 the restriction x0 | ∩ B is extendable to an
L-legal function x1 on B, i.e. the restriction of xi to each translate of D
inside B must be in L.
Remarks (a). Besides the limit of dim Fix 0i /(0/ 0i ) the totality of the numbers
dim Fix 0i for all lattices 0i ⊂ 0 carries an interesting information about (X, 0).
P if 0 = Z and 0i = iZ, this information is encoded in the generating
For example,
function i t i dim Fix 0i which we shall study in the second part of this paper.
(b) One can make the above definition of resdim more robust by using δ-fixed
points Fixδ 0i , i.e. moved by judiciously chosen generators of 0i by at most δ.
Also, one may use Widimε Fix 0i for the metric supγ ∈0 |x − x 0 |γ on Fix 0i instead
of dim Fix 0i , where eventually δ, ε → 0.
1.10.1. Residual Amenability

This signifies the existence of a decreasing sequence of normal subgroups
T 0i ⊂ 0
with amenable quotient groups 1i and with trivial intersections, i 0i = {id}.
Now, for each 1i , we may have some notion of mean (legal, projective, etc.)
dimension which passes to 0 as we apply it 1i acting on Fix 0i and let i → ∞.
Alternatively, one may take an amenable sequence i ⊂ 1i and use Fixδ 0i with
respect to the metric supγ |x − x 0 | for γ ∈ 0 summing over the pull-back of i
under the quotient map 0 → 1i .
In fact, the natural class of groups where this idea works consists of all initially
subamenable groups (essentially introduced in [Ve-Go] and used in [GroESAV ])
generalizing residually amenable groups.
1.11. LINEAR LAWS AND MEAN DIMENSION OVER AMENABLE ALGEBRAS
Given an arbitrary field K, one may take a vector space X over K, e.g. X = K s ,
and speak of linear laws (i.e. subspaces) L ⊂ XD . Then, if 0 is an amenable group,
we have our (mean) projective dimension
prodim(Y : 0) for Y = Y (L) ⊂ X0
defined with an amenable exhaustion of 0.
Remark on finite fields K. If K is finite, then X = X0 is compact (totally dis-

connected) for the product topology and Y ⊂ X is a closed (and so also compact)
subspace in X. Then the basic topological invariant of the action of 0 on Y , the
topological entropy, is (obviously) related to the mean dimension by the equality
topent(Y : 0) = prodim(Y : 0) log |K|,
for |K| =def card K. (See Section 2 for continuation of this discussion.)
Replacing (K s )0 by K s (0) and passing to (group) algebras. Instead of the

space (K s )0 of all functions 0 → K s one can look at the dual space denoted
K s (0) which can be identified with the space of functions with finite support on
0. Then each linear law L ⊂ (K s )D defines a subspace Y0 = Y0 (L) ⊂ K s , namely
Y0 = Y (L) ∩ K s (0) for the obvious embedding K s (0) ⊂ (K s )0 and, clearly,
prodim Y0 = prodim Y . Then we observe that K s (0) can be identified with the free
module of rank s over the group algebra K(0) where the (0-invariant!) subspaces
Y0 ⊂ K s (0) are just submoduli in K s (0).
Now we generalize everything to an arbitrary K-algebra A in place of K(0).
We say A is amenable if it admits an amenable exhaustion by K-linear subspaces
Ai ⊂ A, i = 1, 2, . . ., where amenability of {Ai } signifies that Ai , for large i are
‘almost invariant’ under right multiplication in A, i.e.
(dimK Ai + Ai a)/ dimK Ai −−→ 1
i→∞
for each a ∈ A.
348 MISHA GROMOV
Next, given a finitely generated left module B over A, we define its dimension
relative to {Ai } as follows. Take some finite K-dimensional linear subspace B0 ⊂
B generating B over A and set
dimA B | {Ai } = lim inf dim Ai B0 / dim Ai .
i→∞
Clearly, this dimension does not depend on the choice of B0 and it gives the ‘right’
number for free moduli: dimA As = s for all amenable exhaustions. Furthermore,
if A equals the group ring K(0) of some 0, this reduces to the notion of le-
gal (or stable) dimension over 0, but I do not know if the existence of the limit
limi→∞ Ai B0 / dim Ai holds in full generality.
Remark (made by Ofer Gabber). Since lim inf is nonadditive, we cannot claim
the additivity
dimA B1 ⊕ B2 = dimA B1 ⊕ dimA B2
prior to proving the existence of the limit. Yet we always can take some general-
ized limit (the best here, I think, is an ultralimit) and thus recapture the additivity.
Eventually we shall be interested in additivity of dimA for exact sequences, 0 →
B1 → B → B2 → 0, where some extra problems arise (as was also pointed out to
me by Ofer).
Let us relax the assumption of B being finitely generated over A by giving
B a topology where the action of A is continuous and such that B admits a dense
finitely generated submodule B 0 . (For example, if A = K(0), one can take B equal
the space of all functions 0 → K s with the product topology in this B = (K s )0 ,
where K s comes with the discrete topology. Clearly, the finitely generated module
B 0 = K s (0) densely embeds to this B.) Then we can define dim B as dim B 0 or
(which is essentially equivalent) by approximating the above B0 by some Bε and
taking
lim inf lim Ai Bε / dim Ai .
ε→0 i→∞
The major drawback of all this is the amenability assumption on A. This can
be overcome in the context of the Von Neumann algebras, e.g. for the rings R(0)
for arbitrary countable groups 0. Here K = R and the relevant modules are those
of `2 -functions 0 → Rs as well as their submodules and factor modules (com-
pare 1.12.1. below). The resulting Von Neumann dimension dim`2 B is well defined
for all 0 and if 0 is amenable it equals the above dimA B as an easy argument shows
(explained to me by Alain Connes about 20 years ago and exposed in the case of
`2 -cohomology in [Dod-Mat]).
1.12. VON NEUMANN DIMENSION
Let Y ⊂ (Rs )0`2 ⊂ (Rs )0 be a 0-invariant Hilbert space inside (Rs )0`2 = `2 (0, Rs ),
the space of the square summable functions 0 → Rs . Then for every subset we
define the restriction operator (map) R : Y → (Rs ) `2 for R (x) = x| and let
∗
R : (R )`2 → Y be the adjoint operator. The 0-invariance of Y (trivially) implies
s
that
trace R R∗ card
∗ =
trace R R0
0 card 0
for all nonempty finite subsets and 0 ⊂ 0 (where (Rs )

`2 = (R ) ) and one
s
defines the Von Neumann dimension of Y as
dim`2 (Y : 0) = |−1 ∗
def
i | trace R R (+)
for some (and so for each) finite subset ⊂ 0, where || =def card (see [Con],
[GroAI ] and references therein).
0
Remark. In what follows we use standard embeddings (Rs ) → (Rs ) for all
0
⊃ where we just extend functions by zero outside . In particular, we embed
(Rs ) ⊂ (Rs )0 and observe that R∗ = R0∗ |(Rs ) , and so we abbreviate R∗ to
simple R ∗ for all ⊂ 0.
To see this more geometrically in the case of an amenable group 0 we indicate
the following (well known, I believe)
1.12.1. PROPOSITION. Let i ⊂ 0, i = 1, . . . , be an amenable exhaustion of

0 by finite subsets i and let ni [a, b] denote the number of the eigenvalues of the
operator Ri R ∗ in the interval [a, b]. Then, if 0 < a 6 b < 1,
ni [a, b]/|i | → 0 for i → ∞,
(while ni [0, 1] = s|i |, of course). In other words the majority of eigenvalues is

concentrated near the ends of the α-interval [0, 1].
Proof. Let x: → Rs be an approximate λ-eigenfunction of R R ∗ for some
λ ∈ [0, 1] in the sense that
kR R ∗ (x) − λxk 6 αx (α)
and assume that the restriction of R ∗ (x) to the complement of is β-small, i.e.
kR ∗ (x)|0\k 6 βkxk. (β)0\
We claim that for small α and β the number λ must be close to zero or one. Namely
λ(1 − λ) 6 2α + β. (?)
350 MISHA GROMOV
Indeed, write (β)0\ as

∗
R (x) − R R ∗ (x) 6 βkx|
and obtain with (α),

kR ∗ (x) − λxk 6 (α + β)kαk.
Since kR ∗ k 6 1 and R ∗ R ∗ = R ∗ , we have
kR ∗ (x) − λR ∗ (x)k 6 (α + β)kxk
and

R R ∗ (x) − λRR x (ẋ) 6 (α + β)kxk,
x
i.e.

(1 − λ) R R∗ (x) 6 (α + β)kxk.
Now use (α) again and conclude
λ(1 − λ)kxk 6 (α + β + α(1 − λ))kxk
and, finally,
λ(1 − λ) 6 α + β + α(1 − λ).
In particular we get (?) as well as the relations,
λ(1 − λ) = 0(α + β)
and

α+β
1−λ=0 .
λ
Next let −ρ < be the ρ-interior of , i.e. γ ∈ −ρ iff the ρ-ball B(γ , ρ) ⊂
for a given ρ > 0. We claim that the majority of functions x: −ρ → Rs satisfy
(β)0\ with some β = β(ρ) → 0 for ρ → ∞, at least for finite subsets ⊂ 0.
−ρ
To say it precisely, we denote by Sρ : (Rs ) → (Rs )0\ the operator R0\R ∗ on
−ρ
(Rs ) ⊂ (Rs ) and show that
trace Sρ∗ Sρ 6 β(ρ)|−ρ | (β)ρ
where β(ρ) = β(ρ, 0, s) → 0 for ρ → ∞ and where Sρ∗ : (Rs )0\ → (Rs )0 is
the adjoint to Sρ .
In fact every δ-function x = xγ on 0 concentrated at some γ ∈ 0 satisfies
kR ∗ xγ k ∈ 1 since kR ∗ k 6 1. It follows, that the restriction of kR ∗ xγ k to the
complement of the ball B(γ , ρ) has norm 6 β(ρ) for β(ρ) −−→ 0. Therefore,
ρ→∞
kSρ (xγ )k 6 β(ρ) for all γ ∈ −ρ as well, because 0\ is contained in the
complement of the balls B(γ , ρ) for γ ∈ −ρ . Then the same inequality is clearly
−ρ −ρ
satisfied by Sρ∗ Sρ : (Rs ) → (Rs ) ,
∗
S Sρ (xγ ) 6 β(ρ)
ρ
−ρ
that implies (β)ρ , since the δ-function make an orthonormal basis in (Rs ) .
Now we prove our proposition by first evaluating ni [a, b] for small intervals
[a, b], namely for those where |a − b| = α for some α > 0 specified later on.
We denote by Xi = Xi,a,b ⊂ (Rs )i the span of the λ-eigenfunctions of Ri R ∗
for λ ∈ [a, b] and observe that all x ∈ Xi are α-approximate λ-eigenfunctions for
every λ ∈ [a, b].
Next we consider those x ∈ Xi which vanish on the ρ-boundary of i , i.e.
ρ −ρ
x ∈ Xi =def Xi ∩ (Rs )i and observe that

dim Xρ − dim Xi /|i | −−→ 0
i
i→∞
for every fixed ρ, a and b by the amenability of {i }. Thus the estimate for dim Xi
reduces to that for Xiρ . Then we take the intersection of Xi−ρ with the span of the
eigenfunctions of Sρ∗ Sρ , (with i in place of ) corresponding to the eigenvalues
ρ,β −ρ
6 β 2 . We denote this by Xi ⊂ Xiρ and notice that the operator Sρ : (Rs )i →
ρ,β
(Rs )0\i has norm 6 β on Xi . Furthermore, according to (β)ρ , the dimension of
ρ,β
Xi is rather close to that of Xiρ for large i and ρ. Namely, ∀β > 0, ε > 0∃ρ, s.t.
ρ,β
lim sup dim Xiρ − dim Xi /|| 6 ε.
i→∞
ρ,β
Thus all we need is to estimate the dimension of Xi . To do this we invoke (?),
and apply it to λ = a ∈ [a, b] with a − b = α and get
a(1 − a) 6 2(a − b) + β,
ρ,β
provided some space Xi has positive dimension. In other words, the inequality

|a − b| < a(1 − a) − β /2
ρ,β
forces dim Xi = 0; consequently
ρ
lim sup dim Xi /|i | 6 ε
i→∞
for sufficiently large ρ = ρ(β, ε) and then also

lim sup Xi /|i | = 0
i→∞
since ε → 0 for ρ → ∞. Thus we proved our propositions for all intervals [a, b],
where
|a − b| < a(1 − a)/2, (∗)
352 MISHA GROMOV
since β > 0 can be chosen arbitrarily small. Finally, we cover an arbitrary interval
lying strictly within [0, 1] by those satisfying (∗) and thus conclude the proof. 2
1.12.2. COROLLARY. Let Ei ⊂ (Rs ) i

`2 be the image of the unit `2 -ball in a 0-
invariant Hilbert subspace Y ⊂ (Rs )0`2 under the restriction map Ri . Then
lim(Widimε Bi )/|i | → dim`2 (Y : 0)
for each ε in the interval 0 < ε < 1.
In fact, Ei is a full ellipsoid in the Euclidean space (Rs ) i

`2 where the majority
of the principal semiaxes λν , λ = 1, . . . , |i |, is concentrated at zero or at one.
It follows, the average of λν , eventually defining the Von Neumann dimension
dim`2 (Y : 0), is essentially determined by λ’s close to one and our claim follows
since Widimε of an ellipsoid E with semiaxes λν equals the number of λν ’s greater
than ε 0 for some ε 0 in the interval ε 6 ε 0 6 2ε.
Remark. It is obvious that the number n(ε 0 ) of λν > ε 0 satisfies
n(ε 0 ) > Widimε E for ε 0 = 2ε
while the inequality
n(ε) 6 Widimε E
trivially follows from 1.1.2. Probably, it is not hard to evaluate the critical ε 0 for
which n(ε 0 ) = Widimε E.
1.12.3. The restriction maps Ri : Y → (Rs )i arise from the evaluation map
Re : Y → Rs for y 7→ y(e) for the identity element e ∈ 0. Now, let R: Y → RN
be an arbitrary bounded operator and let Ri : Y → (RN )i be the orthogonal sum
of the γ -translates of R for γ ranging over i . We define Ei ⊂ (RN )i as above
with Ri in place of Ri and let
dR = lim lim Widimε Ei .

ε→0 i→∞
Then a straightforward generalization of the above arguments shows that the supre-
mum of dR over all operators R: Y → RN , N = 1, 2, . . ., equals the Von Neumann
dimension dim`2 (Y : 0).
`p -Remark. The above definition of dim`2 (Y : 0) via dR makes sense for an

arbitrary Banach space Y with a 0-action. Here one can make some modifications,
e.g. by using the `p -norm in the Cartesian power (RN )i for p 6= 2 (compare 1.6.1)
and/or to allow more general (linear and nonlinear) maps R from Y to suitable
spaces. Eventually this line of thought converges to the discussion in 1.6.1.
QUESTION. What is dim`p (Rs )0`q for amenable groups 0? One may (?) expect
that dim`p (Rs )0`p = s for all p in the interval 1 < p < ∞, where the major issue is
the inequality dim`p R0`p < ∞. This would imply, in particular, that the `p -spaces
(Rs1 )0`p and (Rs2 )0`p are not 0-isomorphic for s1 6= s2 , at least for amenable groups
0. This seems to be unknown even for 0 = Z.
1.12.4. Nonlinear Von Neuman

The classical definition of the Von Neumann dimension extends to certain infinite
dimensional smooth manifolds with invariant measures where the tangent bundles
admit 0-invariant Hilbert structures. More generally, let 0 act on a compact space
X with a probability measure µ and let T → X be a Hilbert bundle, such that
the action of 0 lifts to T and preserves the Hilbert norm in the fibers Tx ⊂ T ,
x ∈ X. Then we take the space X of L2 (µ)-sections s: X → T acted upon by the
Von Neumann algebra A generated by all γ ∈ 0 acting on the sections and by the
operators s 7→ f s for all continuous functions f : X → R. With all this, one has
a bona fide Von Neumann dimension dimA X. (If the measure is concentrated at a
single fixed point x0 ∈ X, then dimA X = dim`2 (Tx0 : 0).) And if 0 is trivial the
above becomes the ordinary rank of T , i.e. dim Tx .
Remark. While our mean dimension parallels the topological entropy, the above
Von Neumann dimension is reminiscent of the metric entropy. This may suggest the
following questions. Which (infinite-dimensional) 0-manifolds X have dimA χ 6
dim(X : 0) and when does supµ dimA χ = dim(X : 0) for µ running over
all invariant probability measures on X? However, we do not expect the positive
answer, unless the definitions are modified in some (?) way (compare 2.1).
1.13. TRANSCENDENCE DEGREE FOR 0 - FIELDS
Let F be an extension of a given field K and let 0 act by automorphisms of F

fixing K, i.e. we are given a homomorphism 0 → Gal(F /K).
BASIC EXAMPLE. Consider independent variables xγ associated to all γ ∈ 0

and take F equal the field of rational K-fractions (functions) in these variables. In
other words, F equals the field of rational functions on K 0 viewed as a proalgebraic
variety.
In general, we assume that F is finitely 0-generated over K, i.e. there exists a
subfield F0 ⊂ F whose 0-translates generate F (as the above K(xid ) 0-generates
F = K{xγ }) and then define 0-transcendence degree of F over K with a given
amenable exhaustion {i } of 0 as follows. Let Fi ⊂ F be generated by γ F0 for all
γ ∈ i and a given F0 0-generating F . Then
trandeg(F : {i }) = lim inf |i |−1 trandeg Fi /K.

def
i→∞
354 MISHA GROMOV
We shall not pursue this algebraic line of thinking anymore but shall return to
proalgebraic varieties Y in 2.7 where their dimension will be studied within a
topological framework. (If Y is irreducible then we can pass to the function field F
and define dim(Y : 0) as trandeg(F : 0). But isolating an irreducible component
in (reducible) Y may cause a problem as this may be not 0-invariant.)
1.14. 0 - SPACES WITH SLOWLY GROWING dim Y | i
If our Y ⊂ X0 is given by a balanced (or determined) law L ⊂ XD , i.e. L has

codim L = dim X, then we expect that the spaces L of L-legal functions on
⊂ 0 (see 0.2) have dimensions much smaller than || and this is even more
likely to happen for the overdetermined case where codim L > dim X (here we
deal with discrete groups 0 and finite subsets D ⊂ 0) for ‘sufficiently amenable’
, i.e. having relatively small boundary ∂ρ . For example, if 0 = Zn and i
n−1
are the i-balls in 0, then we expect that dim Li is asymptotic to i n−1 (≈ |i | n )
rather than to i n (≈ |i |), since solutions of balanced difference equation should
be determined by their values on a suitably ‘Cauchy (hyper)surface’ in Zn , e.g., on
Zn−1 ⊂ Zn .
The above suggests a modification of our definitions of various dimensions (X :
{i }) where the cardinality |i | is replaced by |i |β for some β < 1 or by a
more general function α(|i |). Then one can speak of the critical exponent, that
is the maximal (or, rather supremal) β, such that the β-dimension is infinite for
all sequences i ⊂ 0 with |i | → ∞. Next, one may try to compute the βcrit -
dimension with some ‘most amenable’ exhaustion {i } of 0. This will be done
for some examples in the second part of this paper. Here we only observe that for
linear laws L the following three conditions are equivalent:
(1) prodim(Y : {i }) > 0, i.e. lim infi→∞ dim Li /|i | > 0;
(2) lim infi→∞ dim Li /|∂ρ i | = ∞ for all ρ > 0;
(3) there exists a nonzero function y: 0 → X from Y = Y (L) with finite support.
Indeed, obviously, (1) ⇒ (2) ⇒ (3) ⇒ (1) for all amenable sequences {i } in 0.
Notice that (3) says in effect that the implied homomorphism K s (0) → K r (0)
(for K s = X, compare 1.11) is noninjective. Also, one can replace ‘with finite
support’ in (3) by y ∈ `2 (0) in the case K = R.
1.15. MEAN POINCARÉ POLYNOMIAL
Next topological invariant coming after dimension is the Poincaré polynomial of a

metric space X encoding its Betti numbers. This can be modified to Poincarε X
by factoring away ε-fillable classes in H∗ (X), i.e. realizable by cycles C with
FilRad C 6 ε for the metric on C induced from X (compare [GroFRM ]). Thus
the degree of Poincarε X equals the filling dimension mentioned in 1.1.6. Then we
enlarge the metric |x − x 0 | in a 0-space X to |x − x 0 | for ⊂ 0 (see 1.2) and
define the mean Poincaré polynomial as a limit of suitably normalized polynomials

Poincarε Xi , for Xi =def (X, |x − x 0 |i ). Namely, we take
lim lim (Poincaréε Xi )1/|i | . (∗)

ε→0 i→∞
For example, the zero degree term of this limit equals exp(topextent(comp Y :
{i }) where comp Y denotes the space of connected components of Y .
Remark. The above ‘normalization’ by the |i |−1 exponent is motivated by

the exponential bound on the Poincaré polynomial for algebraic laws. Namely, if
L ⊂ (Rk )D is an algebraic variety, then the Poincaré polynomial P (t) of the space
L of L-legal functions on is bounded by exp Ct || as follows from Petrovski-
Thom-Milnor inequality. But the behavior of coefficients of fixed degree (i.e. of
individual Betti numbers) (as well as the convergence in (∗) for i → ∞ with
suitable i ) is a more delicate matter which we do not study in this part of the
paper.
2. Evaluation of the Mean Dimension for Subshifts of Finite Type

We exhibit in this section a variety of examples, where the dimension of a subshift
Y ⊂ X0 equals dim X minus the number of (difference) equations defining Y .
2.1. PRODIM AND LEGDIM IN THE LINEAR CATEGORY
Let X be a finite dimensional vector space over a field K, say X = K s (e.g.,

K = R), and take a subshift Y = Y (L) ⊂ X0 defined by a linear law L ⊂ XD ,
D ⊂ 0 (see 1.8.2). We observe that the projective dimension of such a linear Y
equals its legal dimension, i.e.

prodim Y | {i } = legdim Y : {i } (◦)
for all amenable sequences i ⊂ 0. (See Section 0 and 1.9 for notations.)
Proof. Let L ⊂ X be the space of legal K s -valued functions on , denote
by Mρ ⊂ L the subspace of functions → K s vanishing on the ρ-boundary of
, i.e. on ∩ ∂ρ , and observe that dim Mρ > dim L − card ∂ρ . On the other
hand, if ρ is sufficiently large, i.e. if the ρ-ball in 0 around the identity contains our
D ⊂ 0, then Mρ naturally embeds into Y = Y (L), where each function x: →
K s extends by the identical zero on 0\ and where the inclusion D ⊂ ρ-ball
ensures the L-legality of such extension. Now, if i is an amenable sequence, then
(card i )−1 dim Mρ i and (card i )−1 dim Li have the same asymptotice behavior
for i → ∞ by the above inequality and our claim follows. 2
2.1.1. Remarks. (a) Nonlinearity. If L is a nonlinear law it makes little sense to

make y ∈ Y = Y (L) ⊂ X0 vanish at infinity but instead of this one can look at the
356 MISHA GROMOV
pairs (y1 , y2 ). yi ∈ Y , i = 1, 2, such that y1 (γ ) = y2 (γ ) for all but finitely many

γ ∈ 0. Then one can easily show for many nonlinear subshifts of finite type that
such pairs are abundant in Y × Y if a suitable mean dimension of Y is positive.
(b) Summability. A more interesting generalization concerns the linear case
where X = Rs and Y ⊂ X0 is a weakly closed linear subspace, i.e. it is closed for
the weak convergence in X0 where xi → x iff xi (γ ) → x(γ ) for each γ ∈ 0. Here
one cannot guarantee that the inequality prodim(Y {i }) > 0 implies the existence
of nonzero y ∈ Y with finite support in 0, but one can ensure the existence of a
square summable y: 0 → R in Y . In fact, for every p > 2,

prodim Y`p | {i } = prodim Y | {i } ,
where Y is a weakly closed linear subspace in X0 = (Rs )0 , where Y`p =def Y ∩
`p (0, Rs) and i ⊂ 0 is an amenable sequence.
Idea of the proof. If a linear space of functions on a finite set , e.g. our Y | has
dimension close to cardP , then it contains many sharply localized (concentrated)
functions y where ( w∈ y 2 (w))1/2 is of the order supw∈ |y(w)|. Furthermore,
one can find many, about dim(our space), such y’s, which vanish on a given subset
in provided this subset has relatively small cardinality (such as ∂ρ i in i for
large i). All this follows by simple-minded linear algebra and, when applied to
Y |i , yields in the limit for i → ∞ ‘many’ nonzero functions y ∈ Y`2 and thus
in all Y`p for p > 2. Actually there is the following standard trick of doing this
very quickly. Let Pi denote the normal projection from `2 (0; Rs ) to the space of
functions Y | i extended by zero outside i ⊂ 0. We think of these operators on
`2 (0; Rs ) as matrices indexed by 0 with entries in GLs R, written Pi (γ , γ 0 ), and
observe that
X
dim Y | i = trace Pi (γ , γ ).
γ ∈i
Next we observe that the functions Pi (γ , γ 0 ) on 0 ×0 weakly converge for i → ∞

to some P = P (γ , γ 0 ) which is invariant under the diagonal action of 0 on 0 × 0.
Clearly, the image P (`2 (Y ; Rs )) is contained in Y`2 and its projective dimension
with respect to {i } equals that of Y . In fact, this argument shows that the Von
Neumann dimens ion of Y`2 equals the projective dimension of Y .
Notice that `2 -functions produced by this method appear as normal projections
of δ-functions to Y`2 ,

1 at a given γ ∈ 0,
δ(γ ) =
0 for γ 0 6= γ
and of more general functions ρ with compact supports on 0. Such a projection can
be obtained in certain cases as the limit (t → ∞) of the heat flow which suggests a
possible (not the only one) nonlinear generalization of the `2 -story and which shall
be discussed further in the second part of this paper.
2.1.2. Comparison between dim`1 , dim`2 and dim`∞

The above suggests that the mean dimension of Y`∞ =def Y ∩ `∞ (0; Rs ) minorizes
the Von Neumann dimension of Y`2 . Here `∞ (0; Rs ) denotes the space of bounded
functions on 0 with the sup-norm and the mean dimension refers to that of Y`∞ ∩B 0
for a ball B ⊂ Rs . I do not know if this actually is true but a similar result is valid
with our topological widimε replaced by its linear counterpart, denoted dim− ε A.
This is defined for (centrally symmetric) subsets A in a Banach space L as the
maximal dimension n of a linear subspace L0 ⊂ L, such that the intersection
A ∩ L0 contains the ε-ball in L0 around the origin. Notice that dim− ε A 6 widimε A
according to 1.1.2 but it remains unclear when dim− ε A > widim ε 0 A.
Next, given a 0-invariant subspace X ⊂ `∞ (0; Rs ), we take its intersection

with the unit ball in `∞ (0; Rs ), call this X1 ⊂ X and project it to the spaces
`∞ (i ; Rs ) by just restricting functions from 0 to i ⊂ 0 as we always do. We
look at the images of X1 , call them X1 |i ⊂ `∞ (i ; Rs ), and set
dim`∞ (X : {i }) = lim lim sup dim−

ε (X1 | i )/|i |.
ε→0 i→∞
Now, the argument in 1.12 shows that if the space X1 is weakly closed in `∞ (0; Rs ),
then
dim`2 (X`2 : 0) > dim`∞ (X : {i }) (∗)
for every amenable exhaustion {i } of 0. In particular, if X1 is weakly closed

and dim`∞ (X : {i } > 0 for some amenable exhaustion {i } then X contains a
nonzero `2 -function 0 → Rs .
2.1.3. Remarks. (a) The present condition of weak closeness is by far less
demanding than the one in 2.1. In particular, the above (∗) applies to the spaces
coming from solutions of linear elliptic PDE.
(b) Every `2 -function on 0 is bounded. Furthermore every nonzero x ∈ X ∩
2 (0; R ) gives rise to many functions in X ⊂ `∞ (0 : R ) by taking sums
s s
`P
γ ∈0 cγ γ x for square summable (i.e. `2 ) functions γ 7→ cγ ∈ R on 0. But it
is unclear if
dim`2 (X`2 : 0) > dim`∞ (X : {i }).
It is not even clear what kind of `2 -condition ensures the positivity of dim`∞ and/or
of the mean dimension. On the other hand, if X contains a single nonzero `1 -
function, then dim`∞ (X : {i }) is positive for every amenable exhaustion {i } of
P an x ∈ X ∩ `1 (0; R ), we get lots of bounded functions in X by
s
0. In fact, given
taking sums γ ∈0 eγ λ(x) for bounded functions γ 7→ cγ ∈ R on 0. These suffice
to prove that dim`∞ > 0 and, probably, to show that dimł1 6 dim`∞ for a suitably
defined dimension dim`1 = dim`1 (X`1 : {i }).
358 MISHA GROMOV
0
Squaring `2 -functions. Suppose we are given a bilinear map Rs ⊗ Rs → Rs ,
denoted (x, y) 7→ x • y and observe that so defined product of `2 -function 0 →
0 0
Rs lands in `1 (0, Rs ). Denote by X`22 ⊂ `1 (0, Rs ) the set of the products of all
x, y ∈ X`22 , take the linear span of X`2 and let `∞ (X`22 ) be the closure of this span
0
in `∞ (0 : Rs ). If our product is sufficiently nondegenerate, then

dim`2 X`2 : 0 6= 0 ⇒ dim `∞ X`22 : 0 6= 0.
QUESTION. When does one have the inequality

dim `∞ X`22 : 0 > dim`2 X`2 : 0 ?
For example, is this true for Cs -valued functions with the component-wise
product Cs ⊗ Cs → Cs ?
2.2. GENERICITY AND 0 - TRANSVERSALITY
Denote by ÷ D the set of γ ∈ 0, such that γ D ⊂ . In other words, ÷ D is

the maximal subset − in 0 such that − D ⊂ . Clearly, the cardinality of this
subset ÷ D in 0 satisfies
| ÷ D| 6 || and |i ÷ D|/|i | → 1
for every finite subset D and each amenable sequence i ⊂ 0. (Notice that if 0
has no torsion, then | ÷ D| 6 || − |D| + 1.) It is equally clear that the subspace
L ⊂ X of L-legal functions on (i.e. those x: → X whose restriction to
every translate γ D ⊂ is contained in L) has
codim L 6 | ÷ D| codim L
and so
legdim(Y : {i }) > dim X − codim L (+)
for all amenable i in 0.
-Transversality. Denote by L̃ ⊂ X0 the pull-back of L ⊂ XD under the

restriction map (projection) X0 → XD and say that L is -transversal, for a given
subset ⊂ 0, if the translates γ L̃ ⊂ X0 are all simultaneously transversal for γ
running over ÷ D. This makes sense, strictly speaking, only for finite subsets
⊂ 0; if is infinite this is understood as 0 -transversality for all finite subsets
0 ⊂ .
It is clear that
\ \
L = γ L̃ and Y (L) = γ L̃.
γ ∈÷D γ ∈0
Thus -transversality implies that

codim L = | ÷ D| codim L
and 0-transversality yields the ‘expected’ identity
legdim(Y : {i }) = dim X − codim L (=)
for all amenable i ⊂ 0.
Now we want to decide how generic is the 0-transversality assumption. To do
this we represent L ⊂ XD by r independent linear equations with r = codim L,
i.e. we make L = Ker α for some linear map α: XD → K r , let α: X0 → (K r )0
be the corresponding difference operator and α 0 : K r (0) → K s (0) be the dual
K(0)-morphism for s = dim X (and K s = X). It is obvious that
the morphism α 0 is one-to-one if and only if L is 0-transversal.
Thus the 0-transversality problem and issuing relation (=) reduce to deciding
when α 0 is injective.
2.2.1. It is notationally convenient to interchange r and s and look at maps α from

K r (0) ⊂ (K r )0 to K s (0) ⊂ (K s )0 defined by α ∈ Maps ((K r )D → K s ). We
denote by In ⊂ Maps ((K r )D → K s ), ∈ 0 the subset of those α for which
the kernel of α: K r (0) → K s (0) contains no function x: 0 → X = K s with
support in (where we view elements in K s (0) as K s -valued functions on 0 with
finite supports). Clearly, if is finite, this is a Zariski open subset and if s > r
it is nonempty as it contains an injective α: K r (0) → K s (0) corresponding to
an embedding K r → K s . The subset we really want to understand is In =def In0
corresponding to injective α and this equals intersection of InD over all finite D ⊂
0. We see from the above that this In ⊂ Maps ((K r )D → K s ) = K rs|D| equals the
intersection of a countable family of Zariski open subsets and therefore it is rather
large, at least for uncountable fields K. In fact, it is clear that
if s > r, then every α, whose all sr|D| components are algebraically inde-
pendent over the prime field K0 ⊂ K, gives rise to an injective α, where the
corresponding (dual) L satisfies (=).
2.2.2. Suppose K = R and show that

In ⊂ Rsr|D| contains a nonempty open subset for s > r.
To see this let first s = r and observe that every operator of the form 1 +
ε: Rs (0) → Rs (0) is injective if the sup-norm of ε is < 1, since the equation
ε(x) = x has no nontrivial solution for kεk < 1 (where one may allow nonlinear
operators ε as well). Now, if ε comes from ε: Rs|D| → Rs , then the condition
kε 0 k < 1 is ensured by the inequality kεk < 1, where the norm of ε can be taken
relative to an arbitrary norm on Rs and the corresponding sup-norm on Rs|D| =
(Rs )|D| . This yields our claim for s = r and the case s > r trivially follows.
360 MISHA GROMOV
(a) One can relax the condition kεk < 1 to kεk 6 1 provided the equality
kε(x)k = k(x)k, for x = (x 1 , x 2 , . . . , x |D| ), x i ∈ Rs , is possible only for x 1 =
x 2 = . . . x |D| . This situation arises, for example, for diffusion operators on R(0).
Remarks. (b) The above applies to every normed field K, e.g. to Q with a p-
adic norm. For example, if ε is given by an (s|D| × s) matrix with integer entries
divisible by a prime number, then the corresponding morphism 1 + ε: Qs (0) →
Q(0) is injective.
In fact, 1 + ε remains injective if we replace Q by an arbitrary field of char-
acteristic zero. More generally, let λ be an arbitrary complex number and let all
entries of the above matrix ε be integer polynomials in λ without constant terms
(i.e. divisible by λ). Then the corresponding morphism 1 + ε: Cs (0) → Cs (0) is
injective, provided either λ is transcendental or an algebraic integer which is not
a root of unity.
In fact, the field Q(λ) obviously admits a norm making kεk < 1 under the above
assumptions.
(c) Our (implicite function) argument, shows that an injective morphism α:
K r (0) → K s (0) remains injective under small perturbations if it admits a right
inverse, i.e. a morphism β: K s (0) → K r (0), such that β ◦ α: K r (0) → K r (0)
equals 1. Such β obviously exists (and this was used above) for α induced by
an embedding K r → K s , but it is unclear how typical such invertibility is for
general α. To get a perspective, let 0 = Zn and K = C. Then K(0) = C(Zn )
identifies with the ring of regular functions on the torus (C× )n and morphisms
Cr (Zn ) → Cs (Zn ) become homomorphisms from the trivial vector bundle of
rank r over this torus to such bundle of rank s. Then injectivity of α translates
to injectivity of the vector bundle homomorphism on some fiber, while invertibility
of α amounts to injectivity on all fibers. Thus we see that those α for which α is
injective (i.e. those from In) constitute a nonempty Zariski open subset in Csr|D| for
s > r and all finite subsets D ⊂ Zn , while α corresponding to invertible α have a
similar property only for s > r + n.
(d) If one replaces the space Rs (0) by its dual (Rs )0 , then the corresponding
implicit function argument yields surjectivity of maps 1+ε for suitably contracting
(possibly nonlinear) maps ε: (Rs )0 ←- of finite type (i.e. defined via ε: (Rs )D →
Rs for finite D ⊂ 0). In fact, such an argument yields bounded (and also `p for
p < ∞) solutions to the equation x + ε(x) = y for y ∈ Rs (0) ⊂ `∞ (0; Rs ) ⊂
(Rs )0 and then one uses density of Rs (0) in (Rs )0 . Notice that all this applies
to (K s )0 for all complete normed fields K as well as some subvarieties in these
(K s )0 .
(e) Another class of injective examples is provided by positive selfadjoint oper-
ators Rs (0) ←-. For example, if β: Rs (0) → Rt (0) is injective, then β ∗ β : Rs (0)
is positive selfadjoint and so injective.
2.2.3. Let us give a combinatorial condition on D ⊂ 0 which provides a nonempty

Zariski open subset in In = In0 ⊂ Maps ((K r )D → K s ).
DEFINITION. A collection E of nonempty subsets E ⊂ D is called an extremal

family if for every nonempty finite subset ⊂ 0 there exists γ ∈ 0, such that
(γ ) ∩ D ∈ E.
EXAMPLES. (a) If D ⊂ Zn ⊂ Rn , then the collection of the extremal points of

the convex hull of D makes an extremal family as an obvious argument shows. The
same is true for free groups realized as vertex sets in trees, where the convex hulls
are defined as the spanning subtrees.
Next, for an arbitrary collection E of subsets in D define InE ⊂ Maps((K r )D →
K ) as the set of those α: (K s )D → K r which are injective on the subspaces (K s )E
s
for all E ∈ E, where (K s )E embeds to (K s )D by extending functions by zero on

D\E. Clearly, InE is Zariski open and it is nonempty if supE∈E card E 6 s/r. On
the other hand, if E is an extremal family, then InE ⊂ In0 . In fact, if α(x) = 0,
then the support ⊂ 0 of x: 0 → K r must be empty as trivially follows from
the above definitions. Thus we obtain a nonempty open subset in In0 , provided D
admits an extremal family with sup card E 6 s/r.
EXAMPLE. A group 0 is called D-uncoiled if D admits an extremal system

E with sup card E = 1. We say 0 is uncoiled (traditionally, a unique product
group) if it is D-uncoiled for all finite subsets D ⊂ 0. (Notice that free groups
are uncoiled by the above remark and extensions of uncoiled groups by uncoiled
are, obviously, uncoiled. All this is well known, see [Pass].) For such groups we
have our nonempty Zariski open subset in In for all D, provided s > r (which
extend the solution of the Kaplansky problem for these groups, i.e. nonvanishing
of α for s = r = 1 and α 6= 0, see [Pass]).
Next we want to make examples of injective α for s < r, i.e. we want to embed
K (0) to K s (0) for s < r.
r
DEFINITION. Call a subset D ⊂ 0 tree-like if for every finite subset , there

exist γ and γ 0 6= γ in 0 and w ∈ , such that
γ ∩ D = {γ w} and γ 0 ∩ D = {γ 0 w}.
BASIC EXAMPLE. Let a, b ∈ 0 be freely independent. Then the subset {a, b,

a −1 } ⊂ 0 is tree-like. Indeed this reduces to the free case for 0 = F (a, b), where
everything is clear with w ∈ being an extremal point of the spanning tree (convex
hull) of .
Now, define a subset T2 ⊂ Maps((Rr )D → 0 s ) as follows. For a pair of points
δ, δ 0 ∈ D, consider the subspaces Rrδ , Rrδ 0 ⊂ (Rr )D consisting of functions D → Rr
concentrated at δ and δ 0 correspondingly and identify them with the standard Rr .
Then each map α: (Rr )D → Rs restricted to δ and δ 0 , gives us a map of this Rr to
R2s and we declare α ∈ T2 if this map is injective for every pair of distinct points in
D. Clearly T2 is Zariski open and it is nonempty for 2s > r. On the other hand, our
362 MISHA GROMOV
previous argument shows that if D is star-like, then T2 ⊂ In and so we obtain, in

particular, an injective K(0)-morphism from K 2 (0) to K 1 (0), provided 0 admits
a star-like subset, e.g., if 0 contains a free nonAbelian subgroup.
Remark. If the group 0 is amenable, then an easy argument shows that there is
no injective morphism K r (0) → K s (0) for r > s. On the other hand, such maps
may exist for all nonamenable goups. Also one can ask which nonamenable groups
admit star-like subsets where the picture is unclear, for example, for torsion groups
with sufficiently rare sets of relations. (It is easy to construct a non0-invariant
embedding K 2 (0) → K 1 (0) of finite type for every nonamenable group using a
bounded measure contracting ‘vector field’ or such 0.)
Application to the entropy. Let 0 be amenable and the subset In ⊂ Maps((K r )D

→ K s ) be nonempty Zariski open, e.g., 0 is polycyclic torsion free, D ⊂ 0 is an
arbitrary finite subset and r 6 s. If the field K is finite, this In may be still rather
small but it increases as we pass to finite extensions K• of K. In fact it covers
almost all space K•rs|D| = Maps((K•r )D → K•s ) for large K• , i.e.
card(In(K• ))/ card K•rs|D| → 1
for card K• → ∞.
Next we observe that the topological entropy of the space Y = Y (L(K•0 )) for
a given law L ⊂ (K•)D obviously equals prodim(Y : 0) log card K• . Thus we
obtain many examples of subshifts of finite type where we know what the entropy
is,
topent(Y : 0) = (s − r) log |K• |.
Observe that topent(Y : 0) is notoriously difficult to compute for subshifts of finite

type over 0 noncommensurable to Z and the above algebraic systems constitute the
bulk of available examples (compare [Schm]). We conclude by noticing that the
above applies to uncoiled groups and it seems harder to generate such examples
for groups which contain coils or where uncoilness is unknown.
2.3. STABILITY AND TRANSVERSALITY
A (possibly nonlinear) law L ⊂ XD is called ρ-stable on ⊂ 0 if the legal

extendability of functions x from 0 ⊂ (i.e. of x ∈ (X)0 ) to , i.e. to functions
y ∈ L ⊂ (X) , is equivalent to legal extendability to the ρ-neighbourhood 0 +ρ
of 0 , where we require this property for all 0 in , such that 0 + ρ ⊂
(compare [GroESAV ]). It is easy to see that linear laws L ∈ Grr (K s )D which are
ρ-stable on a finite subset make a constructible subset in Grr (K s )D (i.e. a union
of intersections of Zariski closed and Zariski open subsets). Also, if 0 + 3ρ ⊂ ,
and L is ρ-stable on , then every legal function x on 0 + 3ρ can be modified to
x0 , such that x0 | 0 = x | 0 and such that x0 vanishes outside 0 + 2ρ. Thus,
for ρ > Diam D, one can extend such x0 to a legal function on all of 0 vanishing
outside 0 + 2ρ.
LEMMA. If L is ρ-stable on 10ρ-ball ρ in 0 for ρ > Diam D and L is ρ -

transversal, then it is 0-transversal.
Proof. Suppose L is -transversal and prove it is 1 -transversal for 1 = ∪
{γ1 } and some γ1 ∈ 0\. Denote by L⊥ 1 ⊂ (K )
s 1
the intersection of those Lγ ⊂
(K ) for which γ1 ∈ γ D ∈ 1 and observe that all we need to prove is the
s 1
transversality of L⊥1 with L . Then we take the 2ρ-ball B1 = B(γ1 , 2ρ) around γ1
and project (K s )1 to (K s )B1 ∩1 . We denote by L⊥
1 and L the images of L1 and
⊥
of L under this projection and observe that the transversality of these images is
equivalent to the transversality of the original spaces L and L⊥ 1 . But in the stable
0
case these images are the same for smaller , namely for = ∩ B(γ1 , 5ρ),
where the transversality follows from our assumptions. Thus, the proof follows by
induction on card . 2
COROLLARY. The intersection of the subset Tran0 ⊂ Gr(K s )D with the set of
ρ-stable laws on ρ , say Tran0 ∩Stρ , is Zariski open in Stρ .
QUESTION. Under which assumptions on 0, D, ρ does Stρ ⊂ Grr (K s )D contain

a nonempty Zariski open subset?
2.3.1. Disjoint Transversality

This property is very close to the above stability and it expresses the idea of
noninteraction between subsets in 0 separated by distances > ρ, where the space
Y = L0 ⊂ X = (K s )0 serves as the medium of such intersection. Namely, we
say that the space Y ⊂ X is ρ-disjointly transversal if for every finite system of
finite subsets 1 , . . . , i , . . . , m in 0 with dist(i , j ) > ρ, 1 6 i < j 6 m,
the m subspaces Yi consisting of y ∈ Y satisfying y | i = 0, i = 1, 2, . . . , m,
are transversal in Y . Then disjoint transversality means ρ-disjoint transversality
for some ρ > 0.
It is clear that
stability ⇒ disjoint transversality
(where ‘stability’ means ‘ρ-stability on 0 for some ρ’) and in many cases (e.g. for
groups with bounded asymptotic dimensions, see [GroESAV ]) the disjoint transver-
0
sality implies the existence of a stable sublaw L0 ⊂ LD0 ⊂ (K s )D for some finite
subset D 0 ⊂ 0, such that L00 = Y = L0 . On the other hand, disjoint transversality
of the spaces L for all finite ⊂ 0 (with an obvious modification of the definition
where takes the role of 0) is equivalent to the stability as an easy argument
shows.
364 MISHA GROMOV
2.3.2. Open Problem

The old unsolved question concerns the possible values of prodim(Y : 0). The
above considerations indicate many examples, where this prodim is an integer for
certain (torsion free) groups and it is quite easy to make examples where it takes
rational values for groups with torsion. (See [Lúck] for further discussion of this
problem for the Von Neumann dimension.)
As we have seen above, the integrality of prodim follows from the 0-transversa-
lity of a given presentation (i.e. a law L) of our Y ⊂ (K s )0 , and one may ask for
which 0 every Y ⊂ (K s )0 of finite type admits a presentation (possibly in some
0
(K s )0 for s 0 > s) with 0-transversal L. This can be, probably, expressed with a
suitable Grothendieck group K0 (0) = K0 (K(0)) of finitely generated moduli over
K(0) (or a given amenable algebra A in general). Our prodim should give us a
homomorphism, say d: K0 (0) → R, and we also have a homeomorphism i: Z →
K0 (0) where each s ∈ Z goes to [K s (0)] ∈ K0 (0). Now the basic questions read:
What is the image of d? When does it equal to Z ⊂ R or is contained in Q? What
is there in K0 (0)/i(Z)? Do the subgroups i(Z) and ker d generate K0 (0)?
Apparently, all this is well known for polycyclic and, moreover, for elementary
amenable groups, where K0 (0) tends to be quite small (as was pointed out to me
by Ofer Gabber, also see [Lück]).
Another kind of a transversality question is as follows.
Given submoduli Y ⊂ K s (0) and Y0 ⊂ K s0 (0), can one find a K(0)-morphism
ρ: K s (0) → K s0 (0), such that
prodim(Y ∩ ρ −1 (Y0 )) 6 δ
for a given δ > 0? More specifically, when is this possible with δ = prodim Y −s0 +
prodim Y0 ? Or, even better, when can one find ρ mapping Y 0-transversally (in an
obvious sense) to Y0 ? For example, when does, for a given Y ⊂ K s (0), there exist
ρ: K s (0) → K(0), such that the kernel of ρ is 0-transversal to Y ? Also, observe
that the dimension type invariants of moduli lead to norms on K0 (0) and K0 (0)/iZ
(see [Gro]PCMD ). Finally, notice that the K-theoretic point of view does not do
justice to such moduli as K s (0)/ρK s (0) for embeddings ρ: K s (0) → K s (0)
(describing determined systems of independent difference equations).
2.4. MEAN DIMENSION OF SUB - LINEAR SUBSHIFTS
Take a linear subshift Y0 ⊂ X0 for X = Rs and let B ⊂ Rs be a compact subset

containing the origin in its interior. Then, Y = Y0 ∩ B 0 ⊂ B 0 makes a closed
0-invariant subshift (which can be called ‘sublinear’) in the compact (full shift)
space B 0 where one may speak of our mean dimension dim(Y : {i }).
2.4.1. PROPOSITION. The mean dimension of Y = Y (L) ⊂ B 0 equals the pro-
jective dimension of Y0 ⊂ X0
dim(Y : {i }) = prodim(Y0 | {i })
for all amenable sequences i ⊂ 0. Consequently,

dim(Y : {i }) = legdim(Y0 : {i }).
Proof. The upper bound on dim(Y : {i }) follows from 1.9 and we concentrate
on the lower bound. We observe that the projection p: B 0 → B is distance
decreasing for the metric |x − x 0 | on B 0 and the sup-product metric on B . Then
we assume, without loss of generality, that B equals a small ball in Rs around the
origin and then, by applying 1.1.2, obtain the following inequality for the image
Y | =def p(Y ) ⊂ B with the sup-product metric
Widimε Y > Widimε Y |
where Y = (Y, |x − x 0 | ) as earlier. 2
Now the proposition is reduced to the following
LEMMA. There exists ε = ε(X) > 0, such that

Widimε (B ∩ M) = dim M
for all finite subsets ⊂ 0 and every linear subspace M ⊂ (Rs ) ⊃ B (where,
we use the sup-product metric in B ).
Proof. Everything trivially reduces to B being a ball in Rh around the origin
and then B ∩ M appears as the unit ball with respect to the Banach norm in M
induced from the norm in (Rs ) with the unit ball B . Then 1.1.2 applies and the
proof follows. 2
2.4.2. COROLLARY. If s > r, then the subset of the laws L’s in Grr (Rs )D giving
Y of mean dimension s −r is residual. Furthermore, it always contains a nonempty
open subset and, if 0 is uncoiled, then ‘open’ can be strengthened to ‘Zariski open’.
2.5. ON LOCAL DIMENSION OF 0 - SUBVARIETIES
Let X be an algebraic variety (over some field, e.g. R or C) or an analytic space

(over R or over C) or a smooth manifold (where, more generally, we may allow
stratified spaces, e.g. polyhedra). We are interested in subsets Y ⊂ X = X0 defined
by a law which is a subvariety L ⊂ XD in our category. Here, as earlier, one may
first look at the legal and projective dimensions and then try to prove that the mean
dimension of Y intersected with a bounded (and weakly compact) part B ⊂ X0
equals the projective dimension.
Conjecture. If Y is defined by a generic law L ⊂ XD of codimension r, then
legdim(Y : 0) = prodim(Y : 0) = dim(Y ∩ B : 0) = s − r (=)
for s = dim X and B = B 0 , where B ⊂ X is a sufficiently large compact subset.
366 MISHA GROMOV
Notice that the major difficulty in proving that

dim(Y ∩ B : 0) > prodim(Y : 0) (∗)
stems from the problem of bounding from below Widimε Y ∩ B in the finite-
dimensional case (see 1.1.4). Yet, even without resolving the finite-dimensional
problem from 1.1.4, one expects (∗) in many cases, e.g., for Y = Y (L) where
L ⊂ XD is real analytic.
Another (rather technical) issue, which comes about in the smooth category, is
a possible bad behaviour of the singularities of Y = Y (L) and of the intermediate
finite-dimensional spaces L ⊂ X , ⊂ 0. It is unclear if L can be as bad as
any other closed subset but, in any case, one can rule out major (?) pathologies by
imposing genericity assumptions on L, making all L stratified subsets (and often
just smooth submanifolds) in X (compare 2.7).
2.5.1. Many examples of Y = Y (L) ⊂ XD , where dim(Y : 0) = s − r, e.g., for

X = Rs , can be obtained with the implicit function theorem (see 2.2.2). Namely,
we start with some L0 ⊂ X0 , where the equality dim(Y (L0 ) : 0) = s − r is known
for some reason (e.g. L0 ⊂ (Rs )D is a generic linear law of codimension r) and
then apply a difference operator A: X0 → X0 which is sufficiently close to the
identity, e.g. A = Id +ε: (Rs )0 → (Rs )0 , where the implied ε: (Rs )D → R s
is bounded and has small differential. Then Y = A−1 (Y (L0)) ⊂ X0 will have
the same mean dimension as Y0 by the discussion in 2.2. For example, if we start
with L0 ⊂ (Rs )D of codimension r 6 s represented as the kernel of a linear map
(Rs )D → Rr factoring through a coordinate projection (Rs )D → Rs , then every
small smooth C 1 -perturbation Lε of L0 in (Rs )D gives us Yε = Y (Lε ) ⊂ (Rs )0
with dim(Yε : 0) = s − r.
QUESTION. Let L ⊂ (Rs )D be a smooth submanifold of codimension r con-

taining the origin 0 ∈ (Rs )D and denote by L0 the tangent space T0 (L) of L.
What property of L0 ensures that L is 0-transversal near the origin and thus every
L ⊂ (Rs ) is smooth of codimension r| ÷ D| near the origin, i.e. when in-
tersected with Bε , where Bε ⊂ Rs is a ε-ball with ε > 0 independent of ?
Moreover, we want dim(Y (L) ∩ Bε0 ) : 0 = s − r under a suitable assumption
on L0 . (Apparently, what we need is some kind of uniform 0-transversality of L0
meaning, for example, that the corresponding map Rr (0) → Rs (0) is not only
injective, but is left invertible in the `1 -topology.) Also, we wish a more general
result of this type applicable to an arbitrary (non0-fixed) point y0 ∈ Y (rather than
y = 0), where the corresponding tangent space Ty0 (Y ) is not 0-invariant.
2.6. GLOBAL LOWER BOUNDS ON dim(Y (L) : 0) FOR NONLINEAR LAWS L ⊂ XD
Although we have no general result at the moment for ‘local’ mean dimension it is
possible to obtain some lower bounds for ‘global’ infinite dimensional varieties.
An appealing example is where X is the complex projective space and L is

a complex algebraic subvariety, i.e. L ⊂ (CP m )d , d = card D. We shall show
for such L that the corresponding subspace Y ⊂ (CP m )0 of L-legal functions
0 → CP m (see 2.6.14 and 2.7) has the expected mean dimension
dim(Y : 0) = 2m − codimR L, (∗)
provided L is generic in a suitable sense (see 2.7).
In fact the upper bound on dim(Y : 0) follows from that for the legal dimension
legdim Y 6 dim X − codim L (∗)6
which holds true in all categories whenever one has a reasonable notion of gener-
icity for L (i.e. when L appears as a member of a sufficiently ample family of
subvarieties, see 2.7). On the other hand, the lower bound
dim(Y : 0) > dim X − codim L (∗)>
does not need genericity but rather homological nondegeneracy of L (see below)
which, in the case of X = CP n , is satisfied by all algebraic subvarieties L ⊂
(CP m )d . In fact this nondegeneracy is satisfied for many (e.g., sufficiently mobile)
L ⊂ Xd , where X is any complex projective variety but (see 2.7) it is unclear if
(∗)> holds true for all algebraic L ⊂ Xd .
2.6.1. Homological Lower Bound on the Mean Dimension

Let X be a compact finite-dimensional locally contractible metric space, take a
Cartesian power XM = X × X · · · × X with the sup-product metric and con-
| {z }
M
sider a λ-Lipschitz map from a compact metric space into XM , say α: Y →
XM . We denote by N the maximal integer such that the induced homomorphism
N
HČech (XM ) → HČech
N
(Y0 ) does not vanish.
2.6.2. TOPOLOGICAL LEMMA. There exists a positive ε = ε(X) > 0, such that
the ε 0 -dimension of Y for ε 0 = ε/2 satisfies
Widimε0 (Y ) > N,
where this ε does not depend on M, Y , α and N.
Proof. If Y admits an ε-embedding to a polyhedron P then the cylinder of this
map, say Z ⊃ Y admits a metric extending this of Y , such that dist(z, Y ) 6 ε/2 for
all z ∈ Z (compare 2.5). Next we consider the M projections of XM to X, compose
them with α and observe that the resulting maps Y → X extend to continuous maps
P → X for ε 0 6 ε0 = ε0 (X) > 0. In fact X embeds into some Euclidean space,
say X ⊂ Rn , where we may assume our original metric in X is induced by this
embedding. Since our map from Y to X ⊂ Rn is λ-Lipschitz, it extends to a λδn-
Lipschitz map from Z ⊃ Y to Rn ⊃ X. Now, as all of Z is ε 0 /2-close to Y , for
368 MISHA GROMOV
ε 0 = ε/2, our Z lands δ-close to X ⊂ Rn for δ 6 2ε δn. But the δ-neighbourhood

of X in Rn , for small δ 6 δ0 (X ⊂ Rn ) > 0, retracts to X, since X is locally
contractible. This gives us the desired map Z → X extending Y → X and all
these M maps together extend the original embedding Y ⊂ XM to a continuous
map Z → XM .
Finally, if dim P 6 N −1, the above extension allows a homotopy of the embed-
ding Y ⊂ XM to a map which factors through a (N − 1)-dimensional polyhedron.
This makes the induced homomorphism on H N zero. 2
2.6.3. Homological Dimension

Given Y ⊂ X0 we define its projective homological dimension relative to a given
exhaustion i of 0 by projecting Y to Xi , denoting by Ni the maximal dimension
∗ ∗
where the corresponding cohomological map HČech (Xi ) → HČech (Y ) does not
vanish and then by setting
prohomdim(Y : {i }) = lim inf Ni / card i .

i→∞
Next define a similar stable legal dimension for Y = Y (L) ⊂ X0 coming from
L ⊂ XD by
stlehodim(Y : {i }) = lim inf lim Nij / card i ,

i→∞ j →∞
where Nij denotes the maximal dimension of nonvanishing of the cohomology

homomorphism corresponding to the projection Lij : Lj → Li , j > i.
It follows from the continuity of the Čech cohomology under the projective
limits that
stlehodim = prohomdim
in the category of compact metric spaces. This combines with the above topological
lemma and leads to the following
2.6.4. Practical Lower Bound on the Mean Dimension

If X is as in 2.6.A, then Y = Y (L) ⊂ X0 satisfies
dim(Y : {i }) > stlehodim(Y : {i })
for all compact laws L ⊂ XD .
Proof. All one has to add to the above discussion is the following obvious com-
parison between the sup-product metric in X and |x − x 0 | in X0 : the projection
(X0 , |x − x 0 | ) to (X , sup-metric) is 1-Lipschitz. 2
Remark. The dimension stlehodim (despite the ugly notation) is a computable

quantity and so the above lower bound on the mean dimension is practically useful.
2.6.5. Evaluation of Stlehodim in Manifolds

Given a subspace B in a compact space A we denote by 3(B) ⊂ H ∗ (A) the part
of the cohomology of A which can be represented by Čech cocycles supported
arbitrarily near B, where the cohomology is taken with coefficients in a fixed field
K. Notice that λ ∈ H ∗ (A) belongs to 3(B) iff its restriction to A\B vanishes.
This is obvious. Furthermore, if A is a closed manifold, then 3(B) equals the
Poincaré dual of the image of H∗ (B) in H∗ (A). This is a (small) part of the standard
‘Poincaré duality package’ which is attached to all homology manifolds and also
applies (with some precaution) to general Poincaré duality spaces. Here is another
obvious property of 3(B),
(?) if some α ∈ H ∗ (A) restricts to a trivial class on B ⊂ A, then α ^ λ = 0 for
all λ ∈ 3(B).
This will be used below in the following way. Define corank λ, λ ∈ H ` (A),
as the maximal k, such that λ ^ α 6= 0 for some α ∈ H k (A). Notice that if A
is an n-dimensional manifold (or a general Poincaré duality space), then corank
λ = n − ` for all nonzero λ in H ` (A). More generally, given a map A → C,
define corankC λ by means of those α ∈ H k (A) which come from H k (C). Clearly,
corankC 6 corankA = corank.
Next, set
def def
corankC B = corankC 3(B) = sup corankC λ
λ∈3(B)
for a given B ⊂ A. Clearly, (?) implies that the latter corank bounds from below
the maximal dimension k where the homomorphism H k (C) → H k (B) (induced by
B ⊂ A → C ) does not vanish.
| ↑
Finally we observe that
3(B1 ∩ B2 ) ⊃ 3(B1 ) ^ 3(B2 ) (∩)
for all pairs of compact subsets B1 and B2 in A.

Now we return to our power space X0 and Y = Y (L) ⊂ X0 for a law L ⊂ XD ,
D ⊂ 0. Recall that
\
Y (L) = γ YL , (∩G )
γ ∈0
for YL ⊂ X0 being the pull-back of L under the projection X0 → XD . We denote

by 3∗ ⊂ H ∗ (X0 ) the pull-back of 3(L) ⊂ XD under this projection and we want
to apply (∩) to the infinite intersection (∩0 ).
370 MISHA GROMOV
DEFINITION OF H × . Given a commutative (or, skewcommutative) algebra H

we denote by H × the set of formal finite and infinite products,
Y
× ×
H = h = hi , hi ∈ H,
i
where i may run over an arbitrary index set I . We say (and this is
Qall we care about)
that some such h× ∈ H × does not vanish, written h× 6= 0, if i∈J hi 6= 0 for all
finite subsets J ⊂ I .
We shall apply the above convention to H = H ∗ (X0 ) and denote the corre-
sponding H × by H × (X0 ). Here the most interesting infinite products are of the
form
[
h× = γh
γ ∈0
for some h ∈ H ∗ (X0 ) and we want to decide when such an h× does not vanish.
More specifically, we define 3× (Y ) ⊂ H × (X0 ) as the set of the products
^γ ∈0 γ λγ for all assignments γ 7→ λγ ∈ 3∗ and we introduce the following
DEFINITION. Given a finite subset ⊂ 0, consider an α in the image of

H N (X ) in H N (X0 ) (for the projection X0 → X ) and λ× ∈ 3× (Y ), such that
α ^ λ× 6= 0 and let N be the largest integer where such α and λ× exist. Then set
corank(Y : ) = corank(3× (Y ) : ) = N/ card .

def def
2.6.6. PROPOSITION. The above corank bounds from below the stable legal ho-
mological dimension of Y = Y (L) for compact laws L ⊂ XD ,
stlehodim(Y : {i } > lim inf corank(Y : i ).
i→∞
The proof is clear with the preceding discussion. Also, the following corollary
is now obvious.
2.6.7. If X is a closed manifold (or a general Poincaré duality space) and λ ∈

3(L) ∩ H k (XD ) is a class such that its lift λ to H ∗ (X0 ) satisfies ^γ ∈0 γ λ =
6 0,
then stlehodim(Y : 0) > k and consequently dim(Y : 0) > k.
2.6.8. EXAMPLE: UNTANGLED LAWS AND MONOMIALS. Suppose L0 ⊂

XD is given by d untangled (systems of) equations in the (groups of) variables x i ,
i = 1, . . . , d = card D, namely by fi0 (x i ) = 0, i = 1, . . . , d. In other words,
L0 equals the intersection of d-subsets L0i coming from some L0i ⊂ X via the d
projections XD → X (where each L0i ⊂ X may be given by the equation f 0 (x i ) =
0). Then take some λi ∈ 3(L0i ) and observe that their tensor product (monomial)
λ1 ⊗ λ2 ⊗ · · · ⊗ λd is contained in 3(L0 ).
Denote by λ the cup-product of λ1 , . . . , λd in H ∗ (X) and suppose there exists

α ∈ H k (X) such that λ ^ α 6= 0, i.e. corank λ > k. For example, if X is a closed
manifold (or a general Poincaré duality space) of dimensions and λ ∈ H r (X), then
there always exists such an α ∈ H k (X) for k = s − r.
Next, we consider the γ -translates of the monomial λ1 ⊗ · · · ⊗ λd for all γ ∈ 0
and formally cup-multiply them over 0. The resulting 0-monomial clearly equals
the tensor product of 0 copies (translates) of λ, one λ assigned to each γ ∈ 0.
Denote this 0-product by λ0 and observe that, formally,
λ0 ^ α 0 = (λ ^ α)0 6= 0
for the above α ∈ H × (X).
2.6.9. COROLLARY. Let L ⊂ Xd be homologous to L0 , and so 3(L) contains the

above monomial λ1 ⊗ λ2 ⊗ · · · ⊗ λd . Then dim(Y (L) : 0) > k. In particular, if X is
an s-dimensional manifold and L is homologous to intersection of d cycles coming
from some cycles Li in X (via the d projections XD → X, d = card D), where
Pd
i=1 codim Li = r and the homology class represented by their intersection in
H s−r (X) does not vanish, then
dim(Y (L) : 0) > s − r. (>)
2.6.10. Remarks. (a) This corollary is most powerful if applied to the coho-
mology with finite (e.g. Z/2Z) coefficients where the monomial condition is not
so restrictive. Thus starting with a monomial µ0 in H ∗ (XD ; Z) nondivisible by an
integer p, one gets nonmonomial classes of the form µ0 + pµ0 where the corollary
may apply.
(b) If we work with real coefficients, then the nonvanishing of an (infinite)
integer monomial µ0 obviously yields this for µ0 + 2µ0 for an integer µ0 and all
transcendental 2 ∈ R. Unfortunately it is not useful as the cohomology H ∗ (L) ⊂
H ∗ (XD ) lives over Z but it suggests that nonvanishing of products of the form
^γ ∈0 γ λ and issuing lower bound on the mean dimension are generic phenomena.
This is also confirmed by the examples we study below.
2.6.11. Nonvanishing Products over Uncoiled Groups 0

Let H be a (skew)commutative algebra with unit, e.g., H = H ∗ (X; K) and H =
H ⊗0 (i.e. H equals the tensor product of 0 copies of H , say of H γ = H , where
the basic example is H = H ∗ (X0 )).
Q Nonvanishing problem. Given an h ∈ H . Decide when the formal product

γ ∈0 γ h does not vanish, where we use the obvious action of 0 on H .
Psimplest case, and the only one we address here, is where h is ‘linear’, i.e.
The
h = γ ∈D hγ for hγ ∈ Hγ , where D ⊂ 0 is a finite subset and where all hγ are
372 MISHA GROMOV
N
assumed 6= 0. (This is somewhat opposite to the monomial case, h = γ ∈D hγ we
studied earlier.)
P
2.6.12. PROPOSITION.
Q Let h = δ∈D hδ . If 0 is D-uncoiled (e.g., uncoiled,
see 2.2.3) then γ ∈0 γ h 6= 0.
COROLLARY. Let X be a closed (r − 1)-connected manifold (i.e. its homotopy

groups up to πr−1 (X) vanish) and L ⊂ XD be a cycle of codimension nonhomolo-
gous to zero. Then
dim(Y (L) : 0) > dim X − r
for all uncoiled amenable groups 0, e.g., for all nilpotent and polycyclic groups
without torsion.
Q
Proof. We must show that π = γ ∈ γ h 6= 0 for all finite subsets ⊂ 0. We
proceed by induction on card . Choose ω0 ∈ and δ0 ∈ D such that Q ω0 δ0 = γ0
has a unique solution, let − = \{ω
Q 0 } and assume that π− = γ ∈− γ h 6=
0. Then, our full -product
P π = γ ∈ γ h equals the product of π− by the ω0 -
translate h• of h = hδ0 + δ6=δ0 hδ . This translate can be written as
X
h• = ω 0 h = hγ0 + hγ .
γ 6=γ0
Also, observe that no monomial in π− includes a factor coming from Hγ0 , since
− D does not contain γ0 . Thus π = π− h• = π− ⊗ hγ0 + ε, where no ε-term
includes hγ0 as a factor. Hence, no cancellation is possible and π 6= 0. 2
2.6.13. Remark on the Kaplansky problem. This refers to the following ques-
tion. Let 0 have no torsion. Can then the group ring K(0) have zero divisors? The
above generalizes the standard argument showing there is no zero divisor in K(0)
if 0 is uncoiled (see Section 2.2 and [Pass]).
2.6.14. Positivity and Noncancellation in Complex Manifolds

Suppose we have an ordered (graded skewcommutative) algebra H where the order
is given, by definition, by a subset H + ⊂ H consisting of what we call positive
elements, such that H + is closed under addition and multiplication in H and H + ∩
−H + = {0}.
EXAMPLE. Let H be the real cohomology algebra of CP m , i.e. the algebra of

polynomials in a variable t truncated by the relation t m = 0. Then, nonvanishing
polynomials with positive coefficients define an order in the above sense. Notice
that the integral positive elements in this
H = H ∗ (CP m , R) ⊃ H ∗ (CP m ; Z)
are exactly the Poincaré duals of fundamental classes of complex subvarieties in

CP m .
An order on H induces a natural order on every tensorial power H of H where
+
H is defined as the set of sums of tensor products of positive elements in H .
For example, if H = H ∗ (CP m ) then its tensor power H ⊗d consists of truncated
polynomials in d variables with the obvious notion of positivity. It is not hard to
show that the Poincaré duals of complex subvarieties in (CP m )d are positive in this
sense.
Now, if we look at H = H ⊗0 for a group 0 and take some positive element
h 6= 0 there, (i.e. h ∈ H + \{0} for H + defined
Q with some order in H given by
H + ⊂ H ), then the formal infinite product γ ∈0 γ h does not vanish. This applies,
for example, to the classes in H ∗ (CP m )D , D ⊂ 0, dual to complex subvarieties in
(CP m )D = (CP m )d , d = card D, and lead to the following
COROLLARY. Let L ⊂ (CP m )D be a cycle of codimension r homologous to a

complex algebraic one. Then dim(Y (L) : 0) > 2m − r for all amenable groups 0.
Standard order on H ∗ (X; R). The space of real exterior forms on Cn has a nat-
ural (minimal in some sense) order where positive 2k-forms are defined as positive
combinations of pull-backs of the standard (positive!) volume from an Ck under
nonsingular C-linear maps Cn → Ck . (This is the only GLn C-invariant order on
32R (Cn ) but it seems unclear what are other orders on 32kR (C ) for k > 2.) Observe
n
2k
that our positive form lies√ in the subspace of 3R consisting of the form invariant
under the action z 7→ −1z on C , where they constitute a convex cone with
n
nonempty interior. Forms in the interior are then called strictly positive.
Next, given a complex manifold X, a class h ∈ H ∗ (X; R) is called (strictly)
positive if it can be represented by a form which is (strictly) positive on the tangent
spaces Tx (X) for x ranging over an open dense subset in X. Clearly, this is a bona
fide order on H ∗ (X) in our sense.
CLASSICAL EXAMPLE. Embed X into some CP N and intersect it with a generic

hyperplane. Then the Poincaré dual of this intersection is strictly positive in H 2 (X)
assuming dim X > 0.
This shows that positive elements always exist. Moreover, the Hodge theory
says that every complex cycle can be ‘moved’ to the dual of a positive cocycle.
Namely, let HC ⊂ H ∗ (X; R) be the span of the Poincaré duals of the fundamental
classes of complex subvarieties in X.
THEOREM (See [Gri-Ha]). If X is a complex projective manifold, then the strictly

positive elements constitute a cone with nonempty interior in HC . Thus, for every
h ∈ HC , there exists a strictly positive h+ ∈ H ∗ , such that h + h+ is strictly
positive.
374 MISHA GROMOV
Remark. Our Qinterest in positivity is motivated by the nonvanishing problem for

products π = γ ∈0 γ h for some h ∈ H ∗ (X0 ) = (H ∗ (X))⊗0 which eventually
come from h ∈ H ∗ (XD ), D ⊂ 0. We know that π 6= 0 if h is positive for the order
relation in H ∗ (XD ) = (H ∗ (X))⊗d , d = card D, induced by the above order in
H ∗ (X) associated to the complex structure in X. But the order in H ∗ (XD ) coming
directly from the complex structure in XD usually has more positive elements than
those coming from H ∗ (X) (as some algebraic cycles in Xd do not come from
products of such cycles in X’s. For example, graphs of ‘interesting’ automorphisms
of X give us such cycles in X × X).
QUESTIONS. Let h be positive with respect to the complex structure in XD . Does

then π nonvanish? Let L ⊂ XD be a complex subvariety of real codimension r.
Does the mean dimension of Y = Y (L) satisfy
dim(Y : 0) > dimR X − r?
What can be said about ‘positivity’ of the cohomology classes in X and in XD

which are positive on all algebraic cycles?
2.6.15. Representing Infinite Products by Measures

Q
It is hard to make sense of an infinite product π = γ ∈0 γ h for general h ∈ H ,
but if h is positive for a suitable order on the algebra H this can be done.
EXAMPLE. Let X = CP ∞ and so H = H ∗ (X0 , R) equals the algebra of

polynomials in the variables xγ , γ ∈ 0, with the natural action of 0, and with the
standard notion of positivity. If 0 is finite, then monomials are marked by functions
0 → Z+ indicating the degree of the letter xγ in a given monomial. Thus each real
polynomial becomes a function p: Z0+ → R telling the values of coefficients of a
polynomial at all monomials.
Q look from this angle at∗ the0 product over∞an infinite group 0, h 7→∗ π D=
Next,
π0 = γ ∈0 γ h, where h ∈ H (X ), X = CP is induced from h ∈ H (X )
as earlier. The set of monomials in the polynomials γ h, γ ∈ 0, is given by the
double power set (ZD + ) , that is mapped by the above product over 0 to Z+ , denoted
0 0
t: (Z+ ) → ZQ
D 0
+ . This map sends each 0-family of monomials {mγ ∈ Z+ , γ ∈ 0}
0 D
to the product 0 γ mγ ∈ Z0+ . Next, suppose we have functions pγ : ZD + → R,

γ ∈ D, representing polynomials in xγ , γ ∈ D and we want to multiply them
over 0. To do this we limit ourselves to positive functions on the (countable!) set
ZD+ which are viewed as measures on Z+ . Now we can multiply the measures pγ ,
D
×
where the result, denoted p0 = ×0 pγ , is a measure on (ZD 0
+ ) . Of course, this
measure looks rather unruly unless all pγ are probability measures. And if pγ have
finite total masses they can be normalized to have mass one.
Finally, we push forward the product measure p0× to (Z+ )0 via our map t: (ZD +)
0
→ Z+ and declare this to be our infinite product over 0. Notice, that for pγ = γp0
0
the resulting measure is 0-invariant. Also notice that for X = CP m with m < ∞,
we deal with smaller spaces, namely, the finite set {0, . . . , m − 1}D (instead of ZD
+)
and the Cantor set {0, . . . , m − 1}0 (instead of Z0+ ).
Summing up, we see that the ‘fundamental cohomology class’ of the infinite
intersection
\
γ YL, L ⊂ (CP m )D , D ⊂ 0,
γ ∈0
where L is a complex algebraic subvariety and YL ⊂ (CP m )0 is the full pull-back

of L under the projection X0 → XD , is representable by a 0-invariant probability
measure on the (Cantor) set of maps 0 → {0, . . . , m − 1} (where the ‘probability’
property is achieved with an obvious normalization).
The above generalizes to arbitrary ordered real algebras H , where we have to
deal with vector valued measures. The reader may enjoy persuing this more closely.
QUESTIONS. Is there a deeper relation between the algebra-geometric idea of

positivity on 0-varieties (such as X0 ) and (0-invariant) measures on associated
compact 0-spaces? What is the nature of the space of proalgebraic cycles in such
varieties as X0 and Y (L) ⊂ XL where we keep track of the moduli of the cycles
as well as of their ‘homology classes’ expressed by measures on Z0+ ? Is there a
formalism of this kind associated to the Von Neumann algebra of 0?
2.7. GENERIC LAWS L ⊂ XD AND UPPER BOUNDS ON dim(Y (L) : 0)
We want to extend the results of 2.2 to nonlinear laws L ⊂ XD and show that
generically the mean dimension of Y = Y (L) is bounded by what one may expect,
dim(Y : 0) 6 dim X − codim L.
Intuitively, we think that the γ -translates of the equations defining L remain essen-
tially independent for generic laws L ⊂ XD .
2.7.1. Monomial Laws

Let L be the product of subvarieties Lδ ⊂ Xδ = X, δ ∈ D, where we think of XD
as the Cartesian product of Xδ = X, over δ ∈ D. If Lδ = Xδ for all but a single
δ, e.g. if codim L = 1, then, clearly, the translates of L by γ ∈ 0 are mutually
transversal and thus the legal dimension of Y (L) is bounded by dim X − codim L.
But this may fail in general. Take, for instance, 0 = Z/2Z and L = L1 × L2
where L1 = L2 ⊂ X. Then Y (L) = L and codim Y = codim L instead of the
expected value codim Y = 2 codim L. However, the order of things is recovered if
L1 is transversal to L2 in X. Then, clearly, L1 × L2 is transversal to L2 × L1 and so
the resulting Y (L), being the (transversal!) intersection, (L1 × L2 ) ∩ (L2 × L1 ) =
(L1 × L2 )2 , has right codimension (= 2 codim L).
376 MISHA GROMOV
The above reasoning applies to all groups 0, where mutual transversality of

all Li ⊂ X (trivially, compare 2.2) implies that the legal codimension of Y (L) is
> codim L.
2.7.2. Polynomial Laws

Let L be a union of finitely many monomial laws, L = ∪ki=1 Li . Here a simple
example is where 0 = Z, and L defined in RD for D = {0, 1} by the equation
x0 x1 = 0. A sequence {xi ∈ R}i∈Z belongs to Y = Y (L) if and only if xi xi = 0 for
all i ∈ Z, i.e. out of two consecutive x’s one must be zero. Thus dim(Y : Z) = 1/2
rather than zero. Yet, if we perturbe the equation to x0 (x1 − ε) = 0 for ε 6= 0, then
every sequence {xi } ∈ Y (L) looks like . . . 0, 0, 0, x, ε, ε, ε, . . ., where x is a free
variable and thus dim(Y (L)) = 0. This trivially generalizes to all 0 and polynomial
laws L ⊂ XD , where it yields the expected bound on the dimension of Y (L) ⊂ X0
for generic polynomial laws (where all factors of all irreducible components are
mutually transveral in X or at least meet across subvarieties of proper dimensions).
2.7.3. Polynomial Reduction of Algebraic Laws

Suppose we have an algebraic subvariety L = L0 ⊂ XD whichis included in
algebraic family, say Lε ⊂ XD , ε ∈ E, such that some limit Lε→∞ becomes
polynomial in the above sense, i.e. becomes the union of monomial (i.e. product)
varieties with factors in X, where all these factors are mutually dimensionally
transversal, i.e. all intersections Li ∩ Lj ∩ Lk , etc., have codim L1 ∩ Lj ∩ Lk 6
codim Li + codim Lj + codim Lk . Then, if we work in the category of projective
varieties over an algebraically closed field, we come to the following
CONCLUSION.
legdim(Y (Lε ) : 0) 6 dim X − codim Lε (?)
for generic ε ∈ E, where ‘generic’ means away from a countable union of proper
subvarieties in E.
Indeed, the dimension of our intersection is semicontinuous in ε ∈ E and if it is
small for some (possibly asymptotic) value of ε, then it is generically small.
2.7.4. Examples of ‘Polynomial’ Reduction of Algebraic Cycles

The above reduction works very well if X is a projective variety homogeneous
under an action of a linear reductive group A. For example, X may be a manifold
of flags in Cn (e.g., the Grassmann manifold Grn−k (Cn )) acted upon by SLn Cn .
Then, the Cartesian power of the group, Ad , acts on Xd and every algebraic cycle
L ⊂ Xd is included into the family Lε = aε L, aε ∈ Ad .
LEMMA. There is a degeneration (reduction) of L to a ‘prodynamical cycle’

within this family.
Proof. A generic transformation a ∈ A has isolated fixed points in X as fol-

lows from Thom’s transversality theorem (yielding this property for homogeneous
spaces of all connected Lie groups). It follows in the reductive case that there is
a multiplicative 1-parameter subgroup in A, say C× ⊂ A (we work over C here,
to be specific) which acts on X with isolated fixed points. Then, by the complex
Morse theory, such an action must necessarily have a repulsive fixed point, say
x+ ∈ X, such that the eigenvalues of 2 ∈ C× acting on the tangent space Tx+ (X)
have |eigenvalues| > 1. (This was explained to me by Iiosik Bernstein.)
Now let us apply such an action to one component X of Xd and see what it does
to L. For example, let X = P 1 , d = 2, and L is the diagonal. Our action of C× on
P 1 has two fixed points, x+ and x− and aL obviously converges to (P 1 × x− ) ∪
(x+ × P L ) ⊂ P 1 × P 1 , for z → ∞, a ∈ C× . The same eventually happens to
every L ⊂ Xd where we must apply expanding action along various X-factors of
Xd at some points in L. (To see it clearly, we must order all cycles 6ni Ci in Xd as
follows. First we use 6i ni , i.e. the number of irreducible components counted with
multiplicities, which increase in the course of reduction and which is obviously
bounded. Then, we use the dimensions of projections of Lε to the subproducts
X × ··· × X
| {z }
d−δ
which may only decrease in the course of reduction.

When we arrive at a cycle with a maximal number of components where each
of them has minimal dimensions of projections, then this cycle is ‘polynomial’.)
Now we can use the above (?) and conclude to the inequality
legdim Y (aL) 6 dim X − codim L (?• )
for a generic perturbation of L ⊂ XD = Xd , d = card D by a ∈ Ad . 2
2.7.5. Real Case

The above argument does not work directly over R. For example, the North
Pole – South Pole action of R× on S n may collapse all of L to a single point with
all information irrevocably lost. However, we may pass to the complexification
X(C) = A(C)/A0 (C) where A0 is the isotropy subgroup of some x0 ∈ X and
if X(C) is projective, then our conclusion (including (?• )) applies to X = X(R)
acted upon by A = A(R). For example, this works for the above S n acted upon by
SO(n, 1) as the corresponding subgroup A0 (C) is parabolic in this case. But if you
take S n with the SO(n + 1)-action the complexification trick does not work, but
our conclusion may hold true all the same.
QUESTION. Which (homogeneous) spaces X acted by A satisfy (?• )? What about

Rn acted upon by parallel translations and similarity transformations?
378 MISHA GROMOV
2.7.6. Analytic and Smooth Cases

Since the required genericity of L is essentially an algebraic condition, one expects
(?• ) to be valid for complex (and real) analytic subvarieties L ⊂ XD (which may be
noncompact and/or have boundaries). In fact, the required transversality (express-
ing genericity) concerns the behavior of L ⊂ XD at several points x1 , . . . , xN ∈ L
where the lifts of L to X meet. If we could deform the germs at L by Ad inde-
pendently at these points, we could easily arrive at (?• ) in the analytic category. In
fact, such independence is achieved in the real analytic category if we use the group
A of all real analytic transformations of X, where X is an arbitrary real analytic
manifold, and if we work over C, we may admit L’s ⊂ XD which are images of
holomorphic maps f : L̃ → XD where L̃ is a Stein manifold. Then, by allowing
Lε = fε (L̃) for all holomorphic deformations of f , we again recapture (?• ), at
least in the case of a homogeneous X, by a rather standard argument. Yet, I could
not rigorously prove (?• ) as it stands for complex analytic L ⊂ XD .
Finally, the above should work in the smooth category with A = Diff X where
one, probably, needs some equisingularity lemma in the spirit of Thom (compare
1.3.2.(E1 ) in [GroPDR ]) but I did not check the details (appearing rather straightfor-
ward to a casual eye).
2.7.7. Algebraic Laws L ⊂ XD for Nonhomogeneous X

Start with a projective embedding X ⊂ P M = CP M and then embed Xd →
P N for N = (M + 1)d − 1 in the usual way. (For example, if d = 2, a pair
(x0 , x1 , . . . , xM ), (y0 , y1 , . . . , yM ) goes to ((z00 = x0 y0 , . . . , zij = xi yj , . . . , zN =
xM yM ).) We look at the family LS ⊂ Xd obtained by intersecting Xd ⊂ P N
with a linear (i.e. projective) subspace S ⊂ P N of a given codimension `. Among
these LS there exist ‘most degenerate’ ones which are polynomial in our sense
and satisfy the transversality assumptions of 2.7.1. These come by intersecting
Xd ⊂ P N with ‘tensor products’ of subspaces in P M . (For example, the hyperplane
zij = 0 in P M for N = 2(M + 1) − 1 intersect X2 ⊂ P N across the union
(Xi × X) ∪ (X × Xj ) where Xi denotes the intersection of X with the hyperplane
xi = 0 in P M .) Therefore, generic L = LS ⊂ XD = Xd , d = card D, give rise to
Y = Y (L) with the expected legal and mean dimensions,
dim(Y : 0) = legdim(Y : 0) = dim X − codim L. (??)
In fact the upper bound on legdim follows by the above reduction argument
while the lower bound depends on the homological positivity argument in 2.6.
QUESTION. Does this conclusion (or at least the upper bound on legdim(Y : 0))
remain valid for all projective embeddings Xd ⊂ P N ?
3. Harmonic Maps and Related Spaces

We prove here the results stated in Section 0.6 concerning the mean dimension of
spaces of harmonic maps and of solutions of more general elliptic PDE. We start
with a recollection of the standard properties of linear PDE.
3.1. CAUCHY– GÅRDING INEQUALITY
Consider a homogeneous, uniformly elliptic system of linear PDE imposed on

Rs -valued functions x in the standard unit ball B ⊂ Rn , say Ex = 0. If the co-
efficients of the equations are smooth, then the classical regularity theorem ensures
the smoothness of x. Moreover, all derivatives of x at the origin 0 ∈ B are bounded
in terms of the sup-norm of x on B. Here we are mainly concerned with the first
derivative (differential) Dx where the Cauchy–Gårding inequality reads
kDx(0)k 6 C sup kx(v)k (∗)

v∈B
for some constant C = C(E).

Next suppose E is defined over all Rn , where it is assumed uniformely elliptic
and with all coefficients and their derivatives bounded. Then we apply (∗) to each
unit ball B = B(v, 1) ⊂ Rn and obtain a bound on kDxk everywhere on Rn ,
sup kDx(v)k 6 C sup kx(v)k. (∗∗)

v∈Rn v∈Rn
3.1.1. VANISHING COROLLARY. Let x be a bounded solution x of the system

Ex = 0. If x vanishes on an ε-net 6 ⊂ Rn with ε < C −1 , then x = 0.
Proof. If x|6 = 0, then, obviously, kx(v)k 6 ε supv∈Rn kDx(v)k for all v ∈ V .
This and (∗∗) imply that
sup kx(v)k 6 C −1 ε sup kx(v)k

v∈Rn v∈Rn
and so kx(v)| must vanish if Cε−1 < 1. 2
3.1.2. Denote by XE = XLE ∞ the space of bounded solutions x of the system Ex =

0 and restate the above vanishing result as the following
EMBEDDING PROPERTY. The restriction map R6 : XE → `∞ (6; Rs ) ⊂ (Rs )6

is one-to-one.
In fact the above argument implies that R6 is a topological embedding (i.e.
R6−1 :R6 (XE ) → XE is a bounded operator for the uniform topologies) and that
the intersection of R6 (XE ) with the unit ball in `∞ (0; Rs ) is weakly closed in
`∞ (0; Rs ).
380 MISHA GROMOV
3.1.3. Estimate on the Mean Dimension of XE

Take concentric i-balls B(i) ⊂ Rs of radii i = 1, 2, . . ., and let XE (i) ⊂ XE
consist of maps Rn → B(i) satisfying E. Clearly, all spaces XE (i) are mutually
isomorphic via the maps x 7→ ij −1 x sending XE (j ) → XE (i), and the union of
XE (i) equals XE . Furthermore, these XE (i) are compact spaces and one may speak
of their mean dimensions for actions of lattices 0 on Rn compatible with E. Thus
we set
def
dim(XE : 0) = dim(XE (i) : 0),
where the latter dimension does not depend on i.
Finiteness of dim(XE : 0). Let E be invariant under a lattice 0 acting on Rn .

Then
dim(XE : 0) 6 constE vol Rn / 0.
Proof. Use a 0-invariant net 6 ⊂ Rn and observe that the above embedding be-
comes equivariant and sends XE to (Rds )0 where d denotes the number of elements
from 6 contained in a fundamental domain of 0. Thus
dim(XE : 0) 6 ds. 2
3.2. LINEAR PDE ON RIEMANNIAN MANIFOLDS
Let V be a complete Riemannian manifold and consider an elliptic operator E in

some vector bundle over V . If the ‘coefficients’ of E and its ‘ellipticity’ are uni-
formel y controlled by the Riemannian metric, then (∗∗) generalizes to V , provided
the curvature tensor of V is C 1 -bounded (probably C 0 suffices) on V . (Notice, that
we do not need a lower bound on the injectivity radius of V , but we have it anyway
in our applications where (V , E) is invariant under a cocompact group 0.)
3.2.1. The basic examples of such E are as follows:
(A) The ordinary Laplace operator on V .
(B) The Hodge Laplace operator on differential forms.
(C) Various Dirac operators (where one adds sometimes the spin conditions on
V ).
(D) The ∂-operator, in the case where V is Hermitian.
(E) All of the above twisted with an auxiliary vector bundle E over V , with a
Euclidean connection.
Here the inequality (∗) applies to the coordinate charts in V (or in the unit balls
Bv (1) in the tangent spaces Tv (V ), v ∈ V mapped to V by the exponential maps)
and shows that
sup kDx(v)k 6 C sup kx(v)k, (∗∗)V
v∈V v∈V
where the constant C depends only on the curvature of V , i.e. on supv∈V kKv (V )k,
and on the curvature of the implied vector bundle. In particular, we always have
(∗∗)V with some C < ∞ if (V , E) is acted upon by a cocompact (isometry) group.
Then we have the vanishing corollary and embedding property provided (V , E)
has bounded curvature. Furthermore, if (V , E) is invariant under a cocompact
amenable group 0, then, clearly,
dim(XE : 0) 6 constV ,E vol(V / 0).
3.2.2. Remarks and generalizations. (a) The vanishing corollary trivially extends
to manifolds with unbounded curvature if the density ε of a net 1 is allowed to
depend on v. Essentially, we need ε(v) 6 constn kK(v)k−1/2 , where K incorpo-
rates the curvatures of V and E and their first derivatives if so needed. Similarly,
one may admit unbounded section x with ε(v) ≈ (supv∈Bv kx(v)k)−1 for Bv ⊂ V
being the unit ball around v ∈ V .
In fact, one expects here a more generous density bound on 6 in the spirit of
the first main theorem of the Nevanlinna theory.
(b) The above have an obvious version in a general setting where V is an arbi-
trary metric space and XE is replaced by a subspace Y in the space of bounded maps
x: V → Rs . All one needs is uniform compactness of the restriction operators from
Y to functions on the balls B(v, 1) ⊂ V , for all v ∈ V . Actually, one needs even
less: if X is a linear space of bounded functions x: V → Rs where all x ∈ X with
supv∈V kx(v)k 6 1 are uniformly continuous with a given modulus of continuity
then dim(X : 0) < ∞.
EXAMPLE. Let X : X(λ) be a linear space of functions on a Riemannian manifold

V where each x ∈ Y satisfies supv∈V kDx(v)k 6 λ supv∈V kx(v)k for a given
constant λ. Then dim(X : 0) < ∞ and it may be interesting to find more specific
bounds on this dimension in terms of λ and the geometry of V .
(c) The situation similar to the above example arises in the L2 -framework,
where one studies the L2 -spaces XE (λ) ⊂ L2 (V , Rs ) (or sections X → E, in
general) corresponding to the spectrum of E inside the λ-disk in the complex plane.
Here one knows that the von Neumann dimension dim`2 (XE (λ) : 0) < ∞ for
all 0 (cocompact on V ) and λ < ∞. There are several candidates for the `∞ -
counterpart of this space. For example, one may take the weak closure of the above
XE (λ) in L∞ (X; Rs ). Or one may look at some 0-invariant space Y (λ) of bounded
functions, such that E(Yλ ) ⊂ Yλ and supv∈V kE(y)k 6 λ supv∈V kyk for all y ∈ Y .
One wonders whether dim(Y (λ) : 0) < ∞ for such spaces Y (λ).
(d) Let E be the ordinary Laplace operator 1 on functions V → X. Then one
has the following geometric bound on the von Neumann dimension of the space
X1 (λ) of L2 -functions belonging to the spectrum of 1 below λ. Suppose the Ricci
curvature of V is bounded from below by −1 and let N(ε) denote the minimal
number of ε-balls needed to cover of ε-balls needed to cover the quotient space
V / 0. Then
382 MISHA GROMOV

C1 N(λ−1/2 ) 6 dim`2 X1 (λ) : 0 6 C2 N(λ−1/2 )
where the positive constants C1 and C2 depend only on n = dim V .

This easily follows from the Paul Levy isoperimetric inequality (see Ap. C in
[G-L-P]). Notice in this regard that for connected V / 0 the bound Ricci > −1 im-
plies, by Bishop inequality, the following bound on N(ε) in terms of the diameter
of V / 0,

N(ε) 6 max 1, ε −n exp(n Diam V / 0)
where V / 0 is assumed connected and thus

dim`2 (X1 (λ) : 0) 6 const0n max 1, λn/2 exp(n Diam V / 0) (?)
for connected V / 0 (see [G-L-P]).

Notice that the above inequality is very far from being sharp for infinite groups
0, where the following is well known.
(i) Every L2 -harmonic function on V vanishes (as is true for all connected
complete noncompact manifolds V ), by a standard ‘integration by parts’ argument,
(ii) dimL2 (X1 (λ) : 0) → 0 for λ → 0,
where the rate of convergence depends on 0. For example, dim`2 (X1 (λ) : 0)
vanishes for small λ 6 λ(V ) > 0, if and only if the group 0 is nonamenable.
QUESTION. What are the L∞ -counterparts of the above properties? For example,
does the mean dimension of the space of bounded harmonic functions vanish for
all amenable groups 0? (It is clear that dim− introduced in 2.1.2 does vanish.)
(e) Let E → V be a 0-equivariant Euclidean vector bundle of rank s and let
1E be the (Bochner) Laplace operator on sections V . Then the function ϕE (λ) =
dimL2 (X1E (λ) : 0) is related to the above ϕ(λ) = dimL2 (XD (λ) : 0) by the
following classical
KATO INEQUALITY.
Z ∞ Z ∞
e−λδ ϕE0 (λ) dλ 6 s e−λδ ϕ 0 (λ) dλ
0 0
for all δ > 0, where, observe, the derivatives ϕE0 and ϕ 0 are positive (measures)
since our functions are monotone increasing.
COROLLARY.
ϕE (λ) 6 s eλ ϕ(λ). (+)
(f) If E is a ‘geometric’ selfadjoint operator of second order in E then it is

related to 1E by a Boehner formula E = 1E + BE , where BE is a symmetric
endomorphism of the bundle E. Then one can bound the spectral function of E by
that of 1E and the spectrum of BE . Namely, if all eigenvalues of BE in all fibers of

E are bounded from below by −ρ then, clearly,
ϕE (λ) 6 ϕE (λ + ρ) 6 s eλ+ρ ϕ(λ + ρ). (∗∗)
For example, if E = 1k is the Hodge–Laplace operator on k-forms (where s =
(n/k), n = dim V ), then the above BE is minorized by the so-called curvature
operator R = R(V ) and then (?) and (∗∗) give us a spectral bound on 1k in
terms of R (which includes Ricci) and diam V / 0. This applies, in particular, to the
L2 -Betti number bk (V : 0), that is the Von Neumann dimension of the space of
harmonic L2 -forms on V of rank k,

n
bk (V : 0) 6 exp(nv Diam V / 0) (◦)
k
provided ρ(R) > −1. (This was pointed out by Gallot and Meyer for 0 = {e}
in [Gal-Mey].)
QUESTIONS. (a) Can one improve over the eλ -factor in (+)? (Here one may be
willing or unwilling to bring the curvature of E into play.) Can one bound the mean
dimension of the space of bounded harmonic k-form in the spirit of (◦)?
Notice that a bound similar (◦) (but with a poorer dependence on n) holds
true under less restrictive assumption of the sectional curvatures of V (rather than
R(V )) being bounded from below by −1. This is shown in [GroCDB ] for 0 =
{e} but the argument equally applies to all 0. Furthermore, that argument ap-
plies to the homology Hk (V ; K) for an arbitrary field K and yields a bound on
prodim(Hk (V ; K) : 0) for amenable groups 0.
(b) What is the relation between prodim(Hk (V ; R) : 0) and the mean dimen-
sion of the space of bounded harmonic k-forms on V ? (If one had a full-fledged
Hodge theory for bounded forms one could immediately claim the equality of the
two dimensions.)
3.2.3. Harmonic Functions and the Maximum Principle

Let E satisfy the maximum principle, e.g., E equals the ordinary Laplacian 1 on
functions V → R. We claim that
if a bounded solution x of E vanishes on some net 6 ⊂ V (i.e. an ε-net with some
ε < ∞), then x = 0. Consequently dim(XE : 0) = 0 for every amenable group
cocompactly acting on (V , E).
Proof. Let a bounded solution x of E vanish on some net 6 and take a sequence
of points vi ∈ V , i = 1, . . . , such that kx(vi )k → a = supv∈V kx(v)k for i → ∞.
If V is cocompactly acted by 0, we translate all vi by suitable γi ∈ 0 to a fixed
compact subset V0 ⊂ V and then (after taking a subsequence if necessary) pass
to the limit x∞ = limi→∞ γi x. This x∞ vanishes on some (nonempty!) net, say
6∞ ⊂ V , and kx∞ k achieves its maximum at some point v0 ∈ V0 . Hence x∞ (v) =
384 MISHA GROMOV
x∞ (v0 ) for all v ∈ V and since x∞ |6∞ = 0 this x∞ vanishes everywhere. This
yields the vanishing of x as sup kxk = kx(v0 )k.
Next, forget about 0 and just suppose (V , E) has locally bounded geometry.
Then, instead of translating V , we move ourselves to the points vi and pass to the
(pointed Hausdorff) limit manifold V∞ = limi→∞ (V , vi ) with the limit operator
E∞ on V∞ . Then the maximum principle applies to x∞ on V∞ and the proof
follows. 2
EXAMPLE. If E = 1 and we deal with harmonic functions, then the ‘bounded

local geometry’ refers to a bound on the curvature and the lower bound on the
injectivity radius. In fact, the above argument can be easily carried through with
the assumption |K(V )| 6 const alone, without any bound on the injectivity radius.
(Probably, one needs even less, something like K(V ) > − const or Ricci(V ) >
− const.)
Remarks. (a) The above argument, does not use the linearity of E and applies to
all equations satisfying the maximum principle or the convex hull property. (This
includes harmonic and minimal maps into Riemannian manifolds without focal
points.) On the other hand, when we want to evaluate the dimension dim(X E : 0)
we compare two solutions and the linearity is used in an essential way.
(b) Quantitative maximum principle. The maximum principle can be expressed
as follows.
If the value kx(v0 )k is close to sup kx(v)k, then the ratio x(v)/x(v0 ) is almost
constant on a large ball around v0 .
More precisely, let kx(v0 )k > (1 − ε)kx(v)k, for all v in the R-ball B(v0 , R) ⊂
V around v0 . Then kx(v)k > (1 − δ)kx(v0 )k for all v ∈ B(v, r), where δ and r
depend on R, ε (as well as on (V , E), but not on v0 ) and δ → 0, r → ∞ for ε → 0
and R → ∞.
Notice, that this quantitative maximal principle is equivalent to the previously
used one as an obvious limit argument shows. Also observe that the quantification,
i.e. the dependence of δ and r on ε and R, can be made explicit and rather pre-
cise. For example, one can use in the case of harmonic functions and maps, the
mean value theorem expressing x(v0 ) by a weighted average of x(v) on the R-ball.
(Ultimately, one may appeal to the Harnack inequality.)
(b0 ) Notice, that the function x(v) in question need be only defined on the ball
B(v0 , R), not on all of V . Also the almost constancy conclusion remains valid if
the equation E(x) = 0 is satisfied only approximately, kE(x)k 6 εkxk, where
the norm in question is the sup-norm on B(v0 , R) and where we assume that our
x satisfies the Cauchy–Gårding inequality with the constant C independent of the
above ε.
3.3. EQUATIONS WHERE dim(XE : 0) > 0
The L2 -index theorem provides many instances where dimL2 (XE : 0) does not
vanish but it is unclear if this implies nonvanishing of the mean dimension. On
the other hand, the presence of a nonzero L1 -solution of the equation Ex = 0
(trivially) yields sufficiently many bounded solutions to ensure nonvanishing of
the mean dimension dim(XE : 0). An obvious way to go from L2 to L1 is by
taking ‘squares’ of x’s (compare), but this is usually incompatible with the equation
Ex = 0. A happy exception is the Cauchy Riemann ∂ operator as the square of a
holomorphic function is holomorphic. More generally, if V is a complex manifold
and E → V is a holomorphic vector bundle, then one can take, for instance, the
symmetric square of E, denoted E 2 , and observe that the symmetric square of a
holomorphic section is holomorphic. Thus
dimL2 (XL∂ 2 : 0) > 0 ⇒ dimL∞ (XL∂⊗2

∞
: 0) > 0
(but it is unclear if dimL∞ (XL∂⊗2

∞
: 0) > dimL2 (XL∂ 2 : 0)).
EXAMPLES. (a) Let V = Cm and Eλ → Cn be a line Hermitian holomorphic

bundle, i.e. with a given fiberwise norm) where the curvature equals λ dz dz on Cm
for λ real (where dz dz is the standard Hermitian form on Cm ). This E can be iden-
tified with the trivial bundle Cn × C → Cn , such that the norm of the unit section
x(v) equals exp −λkvk2 . If λ > 0, the unit function x1 : Cn → 1 ∈ C becomes
a holomorphic section on Cn which decays as exp −λkvk2 and so is summable
with all degrees. It easily follows, that the space of bounded holomorphic sections
of Eλ has mean dimension equal cn λn for some constant cn > 0. (Here we refer
to the mean dimension with respect to some amenable exhaustian of Cn . If Eλ is
equivariant with respect to some Lattice 0 ≈ Z2n acting on Cn , then this space, say
Xλ , has dim(Xλ : 0) = cn λn vol(Cn / 0).)
(b) Let E → V be a line bundle equivariant for some cocompact group 0 acting
on V .
If the curvature form ω of E is everywhere greater than the curvature κ of the
canonical bundle, i.e. w − κ is positive definite on V , then the L2 -Euler charac-
teristic of (the sheaf of sections of) E equals the L2 -dimension of H 0 (V , E), i.e.
the space of holomorphic L2 -sections V → E. This is the standard corollary of
the vanishing theorems. On the other hand, the Euler characteristic is given by a
certain characteristic class which is a topological invariant of (E, V ) and which is
of the order c1n (E) for bundles E with large c1m , n = dimC V . Therefore, if w > 0,
this class for E i is about i n for large i and so a sufficiently high power E i admits a
nonzero holomorphic section, provided E is a positive line bundle, i.e. its curvature
from w is positive definite.
Remark. Notice that the above can deliver sections for a given E, without taking
powers, provided w −κ > 0 and χ(V , E) > 0. But if we allow E i , there is no need
to appeal to the L2 -index theorem. In fact a simple application of the L2 -estimate
386 MISHA GROMOV
for the ∂-operator (which is essentially based on the Fredholm alternative, a baby
version of the index theorem) yields lots of L2 -sections of E i without any 0-action
at all.
3.3.2. Recollection on L2 -estimates

Let V = (V , g) be a complete Kähler manifold and E → V a Hermitian line
bundle such that w − κ > λg where, as above, w = w(E) denotes the curvature of
E, κ = κ(V ) stands for the curvature of the canonical line bundle of V and λ > 0
is some real number. Then, for every smooth E-valued (0, 1)-form z with ∂z = 0,
there exists a smooth section y: V → E, such that
∂y = z and kykL2 6 const λ−1 kzkL2 , (∗)
where ‘const’ is universal.
This is a by now standard interpretation of the ∂-estimates (see [Nap] and
references therein).
We shall apply (∗) in order to approximate a given smooth section x0 : V → E
by a holomorphic one as follows. Consider z = ∂x0 , solve ∂y = z and take x =
x0 − y. This x is clearly holomorphic, ∂x = ∂x0 − ∂y = 0 and
kx − x0 kL2 6 const λ−1 k∂x0 kL2 . (∗∗)
This x is close to x0 if k∂x0 kL2 is small and/or λ is large. In what follows we
shall be dealing with a manifold V with bounded curvature and high power E i of
a positive bundle E. Thus we assume λ > const and (∗∗) becomes
kx − x0 kL2 6 k∂x0 kL2 . (∗∗∗)
For example, if we want to have a nonzero holomorphic section x of E, all we
need is an x0 , such that k∂x0 k`2 < kx0 k`2 .
3.3.3. LEMMA (see [Tian]). Let E be a positive line bundle on V , and v0 ∈ V

a given point. Then there exists a sequence of smooth sections xi of E i with the
following properties.
(1) All xi are supported in a given (small) ball B(v0 , ρ) ⊂ V .
(2) All xi are holomorphic in a smaller concentric ball B(v0 , ρ0 ) ⊂ B(v0 , ρ).
(3) kxi (v0 )kEi = 1 and kxi (v)kEi < 1 for v 6= v0 .
(4) kxi kL2 > const i −n for some ‘const’ independent of i.
(5) The pointwise norm of ∂xi exponentially decays for i → ∞, k∂xi (v)kEi 6 α −i
for some α > 1 and all v ∈ V .
Proof. Since E is positive, there obviously exists a local holomorphic section
x0 near v0 with kx0 (v0 )kE = 1 and kx0 (v)k < 1 for v 6= v0 . (Actually such an x0
exists on a rather large neighbourhood of x0 , but this is irrelevant at the moment.)
We smoothly extend this x0 to a smooth section x1 : V → E with a support near
v0 and still having x1 (v) < 1 for all v 6= v1 and finally take xi = x1i . This xi
√ 1
is > 1/2 in the ball of radius ≈ 1/ i since kx0 (v)k > const(dist(v, v0 )) 2 for v
close to v0 and so its L2 -norm is at least const i −n . On the other hand, ∂(x1 (v)) is
different from zero away from v0 where x1 (v) 6 1 − ε and so k∂xi (v)k is bounded
by const i(1 − ε)i−1 as required by (5). 2
3.3.4. Remarks. (a) This construction of approximately holomorphic sections

of ‘sufficiently positive’ bundles was explained to me by Simon Donaldson about
5 years ago who used this idea for producing symplectic hypersurfaces.
(b) The above remains true if instead of the powers E i we take an arbitrary
sequence of line bundles Ei → V , such that the curvatures wi = w(Ei ) grow,
roughly, as iw0 for a fixed positive form w0 .
(c) Notice, we did not use the full positivity of E, but rather positivity at the
point v0 in question.
3.3.5. COROLLARY (Existence of holomorphic L2 - and L1 -sections). Let V =

(V , g) be a complete Hermitian manifold as earlier and E → V an Hermitian line
bundle such that w − κ > λg with λ > 0 and such that w is positive at some point
v0 ∈ V . Then some power E i admits a nonzero holomorphic L2 -section. Also E i
admit nonzero holomorphic L1 -sections for all sufficiently large i.
Proof. The existence of an L2 -section is immediate from the preceding discus-
sion and to turn L1 we split E i = E i1 ⊗ E i2 with large i1 and i2 and observe that
the products of two L2 -sections is L1 . 2
Remarks. (a) The L2 -claim remains valid for every line bundle Ei having the
same positivity as E i . Moreover, the holomorphic sections obtained by the above
argument have a controlled decay at infinity. Indeed, let x0 be a continuous section
with compact support and h be the L2 -nearest holomorphic section, i.e. the normal
projection of x0 to the space of holomorphic L2 -sections. Then y0 = x0 − h is
holomorphic outside some ball, say B(v0 , r) ⊂ V , and it is orthogonal to all
holomorphic L2 -sections. Now, take the function ϕ: V → R+ which equals 1
outside a large concentric R-ball B(v0 , R) ⊃ B(v0 , r), which vanishes on B(v0 , r)
and which equals 1 − (R − dist(v, v0 ))/(R − r) for all v in the annulas between
the two balls. Consider the section y1 = ϕy0 and observe that
Z Z
def
(i) hy1 , y0 i = y1 (v)y0 (v) dv > |y0 (v)|2 dv,
V C(R)
where C(R) ⊂ V denotes the complement V \B(v0 , R), and

(ii) k∂y1 kL2 6 (R − r)−1 ky0 kL2
since |∂(ϕy0 )| = |ϕy0 | 6 (R − r)−1 |y0 |. If our bundle is sufficiently positive, we
can approximate y1 by a holomorphic L2 -section y, such that
ky − y1 kL2 6 const k∂y1 kL2 6 const(R − r)−1 ky0 kL2 .
388 MISHA GROMOV
It follows that
Z
0 = hy, y0 i > |y0 (v)|2 dv − const(R − r)−1 ky0 kL2 ,
C(R)
and so
Z
|y0 (v)|2 dv 6 const R −1 |y0 |L2
C(R)
for large R and C(R) = V \B(R).

Finally, as h equals y0 outside B(v0 , R), our h also has its L2 -norm decaying
with the rate R −1/2 at infinity.
(b) Instead of the L2 -nearest h one could take the Lp -nearest one, which is
unique (if it exists) for all p (including p = 1, where the strict convexity is due to
holomorphicity). It seems not hard to show that the Lp -norm of this h has a similar
decay over C(R) for R → ∞.
(b0 ) Let us indicate the proof of the decay property for (as well as the existence
of) holomorphic Lp -sections in the case of locally bounded geometry. First we pass
i
P odd power E of E where one has many holomorphic
to a large L1 -sections (of the
form j xj yj for holomorphic L2 -sections xj of E i1 and yj of E i2 with i1 +i2 = i,
compare 3.3.5). Such E i admits n+1 = dimC V +1 bounded holomorphic sections
x0 , x1 , . . . , xn that are uniformly transversal to the zero 0 = V ⊂ E i and such that
their zeros xk−1 (0) ⊂ V , k = 0, 1, . . . , n, are simultaneously uniformly transversal
(see 4.3). Denote by V e → V the canonical ramified cover of order 2n with the
S
ramification locus 6 = k xk−1 (0), observe that V e is nonsingular and that the lifted
e e e
bundle E → V admits a square root, since E does and i is odd. Now holomorphic
i
L2 -sections X of such square root, say E → V e, can be multiplicatively pushed

forward to holomorphic L1 -sections x of E for x(v) = x(ṽ1 ) ⊗ x(ṽ2 ) · · · ⊗ x(ṽ2n )
for the pullbacks ṽ1 , . . . , ṽ2n of v and so the L1 -properties of E reduce to the L2 -
theory of E. Notice, that the curvature of E (as well as that of E) e vanishes along
e
6 , but only in transversal directions, and so the metric on E can be perturbed to a
one with sufficiently positive curvature, provided we had enough positivity in E to
ensure that E|6 is more positive than the canonical bundles of the submanifolds
x −1
k (0) and their intersections (compare 4.3).
Then we get lots of L1 (and hence Lp , p > 1) of holomorphic section of E with
controlled L1 -decay at infinity.
3.3.6. Let (V , E) be acted upon by an amenable Lie group 0 with V / 0 compact.

Then the space of bounded holomorphic sections of E i for large i has positive
mean dimension.
Indeed, the presence of a single nonzero L1 -section suffices as was mentioned

earlier. (See 3.3.11 for a sharper result.)
3.3.7. Gårding Inequality in E i

In order to see how the Gårding constant for holomorphic sections V → E i
depends on i, we scale the underlying manifold (V , g) by g 7→ ig. Then the
curvature iw of E i scales to w and so we have a uniform (independent of i) Gårding
inequality in E i over (V , ig). Then, coming back to g, we conclude that
The sup-norm of a holomorphic section x: V → E i in the ε-ball B = B(v, ε) ⊂
V bounds the differential of x, by
kDx(v)kEi 6 Cv ε −1 sup kx(v)kEi (+)
v∈B
for every ε 6 i −1 .
3.3.8. COROLLARY. The sup-norm is bounded by the L2 -norm,

Z 1/2
0 −n
kx(v)kEi 6 Cv ε kx(v)k dv
2
6 Cv0 ε −n kxkL2 (++)
B
for ε 6 i −1 .
3.3.9. Remarks. (a) Notice that (+) and (++) are local properties where the
holomorphicity of x is only required on the ball B. Thus we can apply (++) to
the solutions y1 of the ∂-problem ∂yi = ∂xi satisfying the basic L2 -estimate (∗)
from 3.3.2. These yi are holomorphic (as well as xi ) in a small (but fixed!) ball
B(v0 , δ) ⊂ V and then (++) applies to smaller ε-balls B(v, ε) ⊂ B(v0 , δ). It
follows, that the holomorphic sections xi0 = xi − yi converge to xi uniformly (and
exponentially fast for i → ∞) on every concentric ball B(v0 , δ 0 < δ). In fact,
such convergence takes place also on larger balls, where ∂xi 6= 0 anymore, since
the Gårding inequality remains valid for nonhomogeneous situation, but we do not
need this for our purposes.
(b) The constants Cv and Cv0 depend on local geometry of V and E near v.
Actually Cv can be bounded in terms of the curvatures of V and E while Cv0 also
depends on the injectivity radius of V . (In general, ε −n in (++) must be replaced
by (Vol B(v, e0 ))−1/2 for some ε 0 somewhat smaller than ε.) In particular, Cv and Cv0
are bounded if V and E have bounded local geometry, e.g. if there is a cocompact
isometry group 0 acting on V and on E.
3.3.10. INTERPOLATION THEOREM. Let V and E have bounded local geom-

etry and thus the constants Cv and Cv0 are bounded on V , and let 6 ⊂ V be a
δ-separated subset, i.e. dist(σ1 , σ2 ) > δ for all σ1 6= σ2 in 6. Then, for every
i > constV ,E max(1, δ −2 ) and every bounded section y of E i |6, there exists a
bounded holomorphic section x: V → E i , such that x|6 = y.
Proof. First we observe that by scaling the metric g of V , by g 7→ δ −2 g, we
make a δ-separated set 1-separated. This also normalizes the curvature of E i with
i ≈ δ −2 to the unit size and explains (actually proves) the dependence of i on δ.
390 MISHA GROMOV
Now we prove the theorem for δ = 1 by summing up L1 - sections of E i . It

(obviously) suffices for our purpose to have holomorphic L1 -sections xσ• : V → E i ,
for all σ ∈ 6 and a given i > constV ,E , such that
(a) kxσ• (σ )kEi > 1/2;
(b) the sum of the norms kxσ• (σ 0 )kEi over all σ 0 ∈ 6 is small, say 6 0.1.
We recall that L1 -sections are obtained as products of L2 sections and so we need
L2 -sections, say xσ0 , satisfying (a), where (b) is replaced by a similar bound on
the sums of kxσ (σ 0 )k2Ei . Such an xσ is constructed by first using 3.3.3 at v0 = σ
with ρ 6 0.1 and then by approximating the resulting almost holomorphic section,
call it now xσ0 , by a holomorphic one, that is our xσ . The bounded geometry as-
sumption makes the estimates in 3.3.3 independent of σ and then 3.3.8 applied to
ε-balls around all σ 0 6= σ in 6 yield the required bound on the sum of kxσ (σ 0 )k2Ei ,
provided i is sufficiently large. 2
3.3.11. COROLLARY. If (V , E) is acted upon by an amenable group 0 with com-

pact quotient, then the mean dimension of the space of bounded holomorphic
sections of E i is about i n , n = dimC V .
3.3.12. Remark. There is a distinguished holomorphic L2 -section of E i taking

a given value e ∈ Eui at a given point u ∈ V , namely the one which has the minimal
L2 -norm. This section, call it he (v), v ∈ V , controllably decays at infinity in the
sense that the integrals of khe (v)k2 over the complements C(R) ⊂ V of the large
R-balls B(R, u) ⊂ V around u satisfy
Z
khe (v)k2 dv 6 const R −1 .
C(R)
This follows from Remark (a) in 3.3.5 and the Gårding inequality.
3.3.13. Interpolation with Jets and Transversality Theorem

One can easily interpolate not only the values on 6 but also a given number r of
derivatives at all σ ∈ 6. This is done again by first constructing approximately
holomorphic sections and then making them holomorphic by small perturbations,
where ‘small’ refers to the C r -topology as is allowed by 3.3.7 (which needs an
obvious generalization in the case r > 2).
Let us spell out how the approximate sections come about. Start with x0 near
v0 as in the proof of 3.3.3 and let 8 be a finite collection of holomorphic functions
ϕ defined on V near v0 , such that the r-jets of the functions ϕ ∈ 8 at v0 linearly
span the full space of r-jets (as do the set of monomials of degrees 6 r in local
coordinates).
Now we take some sufficiently small positive ε and let xϕ = (1 + εϕ)x0 .
Since ε is small, all xϕ satisfy kxϕ (v)kE 6 1 on the boundary of some small
ball B(v0 , ρ0 ) ⊂ V and so can be smoothly extended with this property to V .

The totality of these extended xϕ represent all r-jets at v0 . This property passes to
the corresponding sections xϕi of E i and then further to the holomorphic sections
approximating xϕi . This is straightforward and left to the reader (who is referred
to [Tian] for further results and applications).
3.3.14. Take a subset S in the jet bundle J r (E i ) over V and let us try to move a
given holomorphic section V → E i away from S. This presupposes some metric
on J r (E i ) and ‘away from S’ means that the r-jet V → J r (E i ) does not intersect
an ε-neighbourhood of S for some ε > 0. In what follows, we assume that V and E
have bounded local geometry and observe that then J r (E i ) also admits a Hermitian
structure of bounded local geometry compatible with this in V . We choose and fix
such structure in each J r (E i ).
We say that S is uniformly k-dimensional, if for each unit ball B ⊂ J r (E i ) and
every δ > 0, the intersection S ∩ B can be covered by at most Cδ −k balls of radius
δ for some constant C = C(S).
UNIFORM TRANSVERSALITY THEOREM. Let E be positive and S uniformly

k-dimensional for k < dim V . Then there exists i0 = i0 (V , E, r), such that for each
i > i0 every bounded section V → E can be moved away from S by an arbitrarily
small (in the uniform topology) perturbation.
Proof. The required perturbation exists over each ρ-ball in V for a fixed small
ρ > 0 as follows from the above and the standard transversality argument. S Fur-
thermore, this argument applies to a union of such balls, say to U = B, µ =
1, 2, . . ., if these balls are situated sufficiently far apart in V . Finally, we cover V
by finitely many U ’s of the above kind, V = U1 ∪ U2 ∪ · · · ∪ UN , and apply
the first perturbation over U1 , then the second, much smaller one over U2 and so
on. This ‘much smaller’ guarantees we do not each step what we gained at the
previous one and so the N’th perturbation gives us a section x: V → E i with the
jet J r (x): V → J r (E i ) missing S, i.e. ‘moves the original section away from S’. 2
3.3.15. Further Applications, Generalizations and Open Questions

(a) As we have mentioned several times earlier, the L2 -part of our discussion ap-
plies to (nonpower!) line bundles Ei with curvature ≈ iω, but to go to L1 (and
thus L∞ ) we need such an Ei to be tensor product of two bundles with this kind of
curvature. Such decomposition is possible, for example, if H 2 (V ; Z) = 0 (but the
interpolation theorem, probably, remains true in all cases, compare 3.3.4(b0 )).
(b) The full L2 -story extends to suitably positive vector bundles E of higher
rank. But our ‘squaring argument’ needs passing to tensorial powers of E. Here
again, it would be nice to prove an L1 -version of the ∂-estimate and this looks
easy.
(c) The proof of 3.3.10 yields on interpolation results for holomorphic Lp -
sections of E i for all p > 1.
392 MISHA GROMOV
(d) The classical correspondence between divisors and line bundles extends
to the framework of bounded geometry. This allows, in particular, construction
of many bounded sections V → E i vanishing on a hypersurface W ⊂ V with
(sufficiently) bounded geometry.
(e) Here are several problems which seem to be solvable in the present frame-
work:
(1) Extension of bounded (and Lp ) holomorphic sections of E i from a submani-
fold 6 ⊂ V of (sufficiently, depending on i) bounded local geometry (where
the case dim 6 = 0 is covered by the interpolation theorem).
(2) Decomposition of bounded holomorphic sections of E i into convergent sums
of L1 -sections.
(3) Construction of bounded sections of affine subbundles P of sufficiently positive
vector bundles. For example, solution of the equation sr=1 cr ⊗ xr = a for
given bounded sections cr of E j and of E i with the unknown xr bounded
sections of E i . Similarly, one is interested in the equation
X
xr ⊗ yr = a
r
where a is an L2 -section and the solution (xr , yr ) must be L1 .

(f) Kodaira embedding theorem. The uniform transversality theorem trivially
implies that the canonical map 2 from V to the projectivized space of holomorphic
L2 -sections V → E i is a holomorphic embedding for i > i0 . (Recall, that 2
is defined by sending each v ∈ V to the space of holomorphic sections of E i
vanishing at v.) Actually, 2 is easily seen to be locally bi-Lipschitz, i.e. there exists
a constant C > 0 such that
C −1 dist(2(v1 ), 2(v2 )) 6 dist(v1 , v2 ) 6 C dist(2(v1 ), 2(v2 ))
for all pairs of disjoint points v1 and v2 in V satisfying dist(v1 , v2 ) 6 1.

If V is compact, then the receiving projective space is finite-dimensional and it
is infinite-dimensional otherwise. In the latter case, we clearly have
dist(2(v1 ), 2(v2 )) → π/2 for dist(v1 , v2 ) → ∞.
There is (apparently) no good finite-dimensional reduction of this map but nice

maps V → CP N are available for all N > dim V within the L∞ -framework.
(g) Many naturally arising line bundles, e.g. those associated to divisors in V
(say with uniformly bounded volumes in the unit balls in V have singular curva-
tures and it would be useful to extend our upper and lower bounds on the spaces of
holomorphic sections to such bundles.
(h) Let E be some Dirac operator twisted with a Euclidean vector bundle E on
V . When can one guarantee the existence of many Lp -solutions to the equation
EE x = 0 (where the cases p = 1 and p = ∞ are especially interesting in the
present context)? Here one exercises a good control over L2 -sections especially
for the tensorial powers E ⊗i in terms of the index of the twisted operator EE⊗i
but it is unclear when there are nontrivial Lp -sections x of E ⊗i satisfying the
equation EE⊗ i (x) = 0. Similar question arises for the Hodge–Laplace operator
acting on 3∗ (V ) where nonzero harmonic L1 -form may (?) appear in the presence
of a nontrivial cup-product.
3.4. NONLINEAR EQUATIONS
Let V be, as earlier, a complete Riemannian manifold and X be a compact Rie-

mannian manifold. We are interested in smooth maps x: V → X satisfying some
elliptic system E of partial differential equations, where basic examples are:
(i) harmonic maps;
(ii) holomorphic maps, where the Riemannian metrics in V and X are assumed
Hermitian;
(iii) maps x: V → X whose graphs Gx ⊂ V × X are minimal subvarieties.
The essential features of our equations we shall need later on are as follows:
(a) Regularity and compactness. Every C 1 -map x: V → X satisfying E is in
fact C ∞ -smooth. Moreover, all higher derivatives of x are bounded in terms of the
first derivatives, i.e.
kD i xk 6 Ci (kDxk) (+)
for some bounded functions Ci = Ci (V , X, E), where k k denotes the sup-norm

on functions on V , i.e. kD i xk =def supv∈V kD i x(v)k. It follows, that the space
of our maps x with kDxk 6 const is compact for the uniform convergence on
compact subsets in V .
(b) Nonlinear Cauchy–Gårding inequality. Let x1 , x2 : V → X be smooth maps,
where x1 (v) can be joined by a unique minimizing with x2 (v) geodesic in X for all
v ∈ V . Then we can compare the differentials
Dx1 (v): Tv (V ) → Tx1 (v)(X)
and
Dx2 (v): Tv (V ) → Vx2 (v) (X)
using the parallel transport in X along the geodesic [x1 (v), x2 (v)] ⊂ X and take
the difference Dx1 (v) − Dx2 (v). Thus we can speak of the C 1 -distance kDx1 (v) −
Dx2 (v)k and set
def
kDx1 − Dx2 k = sup kDx1 (v) − Dx2 (v)k.
v∈V
394 MISHA GROMOV
Notice that this C 1 -distance is well defined if x1 and x2 are C 0 -close, i.e.
def
kx1 − x2 k = sup distX (x1 (v), x2 (v)) 6 ε0 < InjRad X,
v∈V
where, observe, the injectivity radius of X is positive as we assume X is compact.

(Notice that one could equivalently define a C 1 -distance with a given covering of
X by coordinate charts where it is possible to speak of x1 − x2 locable in every
chart.) Now we can state our inequality.
If x1 and x2 have bounded differentials and kx1 − x2 k 6 ε0 for the above ε0 ,
then
kDx1 − Dx2 k 6 Ckx1 − x2 k (∗)
for some constant
C = C(V , X, E, kDx1 k, kDx2 k, ε0 ).
About the proof of (a) and (b) for our examples. The property (a) is well-known
for the classes of maps indicated in the above (i)–(iii) where it is derived from
the corresponding elliptic regularity for nonhomogeneous linear equations via the
standard implicit function argument. The sufficient condition on V and X is a
uniform C 1 -bound on their curvatures. Then (b) follows by the trivial interpolation
property of smooth maps,
kDx1 − Dx2 k 6 Ckx1 − x2 k
for C = C(V , X, kD 2 xk, kD 2 xk).
3.4.1. Embedding Property

Let V and X be as earlier where we assume kK(V )k 6 const < ∞. Consider the
space Xc of maps x: V → X satisfying one of the elliptic conditions (i), (ii) or
(iii) and having kDxk 6 c for a given c > 0. Then there exists ε > 0 depending on
V , X, E and e, such that the restriction map from Xc to X6 for an arbitrary ε-net
6 ⊂ V is an embedding.
This follows from the Cauchy–Garding inequality by the same (obvious) argu-
ment we used in the linear case. Also, we have as a corollary, the bound dim(Xc :
0) < ∞, whenever V is isometrically and co-compactly acted upon by an amenable
group 0 (which must preserve the implied complex structure in the case (iii)).
3.4.2. Dependence of C and ε on c = sup kDxk and the Proof of 0.6.1

Harmonic and holomorphic maps are invariant under the scaling: if x: V → X is
a harmonic (holomorphic) map then it remains such if we replace V by λV and X
by µX, where the notation ‘λV ’ refers to multiplying the metric in V by a constant
λ > 0 and µX has similar meaning. Also observe that the (ellipticity) constant C
in (∗) can be assumed independent of λ and µ in-so-far as these λ and µ are > 1,
since such scaling diminishes the curvature. On the other hand, when we scale the
metrics, the norms of the differentials of the maps x: V → X scale by the rule,
kDxkλ,µ = λ−1 µkDxk,
where kDxkλ,µ is the norm measured with respect to metrics in λV and µX. It fol-
lows, that the constant C in (∗) is bounded by c const(V , X, E, ε0 ) if c = sup kDxk
is > 1. This is seen by taking λ = c and µ = 1. Consequently, the above ε is
bounded from below by δc−1 , δ > 0, and so we obtain the bound of the mean
dimension of Xc for large c by bcn as was stated in (?)∞ of 0.6.1.
Next let us see what happens if c = sup kD(x)k is small. Such a map x sends
large R-balls in V to small ones, of radii cR in X, and if we scale these small balls
to the unit size by passing to µX with µ = (cR)−1 we get maps from B(R) ⊂ V
to almost Euclidean unit balls, where we assume that c is much smaller than R −1 .
Thus, we can think of the harmonic equation for map B(R) → µX on each B(R)
as a small perturbation of the ordinary Laplace equation for maps B(R) → RN ,
N = dim X. Namely, if x1 and x2 are two harmonic maps from B(R) to a unit
ball in µX, then the difference x1 − x2 is approximately harmonic in the Euclidean
sense, where the difference is taken in the Euclidean geometry approximating the
Riemannian one in µX. Now (b) and (b0 ) from 3.2.3 imply the following
APPROXIMATE MAXIMUM PRINCIPLE. Let V and X have bounded local

geometries and let x1 , x2 : V → X be nonequal harmonic maps with kDxi k 6 c,
i = 1, 2, and with kx1 − x2 k 6 ε. Then there is a ball B(v0 , R) ⊂ V where
x1 (v0 ) 6= x2 (v0 ) and the ratio kx1 (v)−x2 (v)k/kx1 (v0 )−x2 (v0 )k is almost constant
on B(v0 , R), where R → ∞ for c, ε → 0 and where ‘almost’ means up to a
(1 + δ)-factor where δ → 0 with c, ε → 0.
This trivially implies (?)◦ in 0.6.A exactly as in the linear case considered
in 3.2.3.
Remarks. (a) We treated above only harmonic maps, but the same argument ap-
plies to the pseudo-holomorphic maps between almost complex manifolds (where
it somewhat simplifies in the honestly holomorphic case).
(b) It is not hard to quantify the above and give a specific bound on dim(Xc : 0)
for harmonic maps and small c in terms of c, the upper bound on the sectional
curvature of X and the rate of decay of the heat kernel in V .
3.4.3. Additional Remarks and Generalizations

(a) One can allow a noncompact target manifold X, provided it has a uniformly
bounded local geometry, i.e. |K(X)| 6 const < ∞ and InjRad X > ε > 0 (where
only the upper bound R(X) 6 const is essential for harmonic maps). Furthermore,
one may start with a general fibration Z → V (instead of the trivial one X × V →
396 MISHA GROMOV
V ) and extend the discussion to sections V → Z satisfying our kind system of

PDE. For example, one has dim(Xc : 0) < ∞ for holomorphic sections of suitable
holomorphic bundles over V , e.g., those associated to the tangent bundle.
(b) If one deals with higher-order elliptic systems one may need a bound on
kD i xk for i > 1 to achieve the full regularity and compactness, where Xc is defined
by the condition kD i (x)k 6 c for some sufficiently large i.
(c) It is interesting to have a possibly precise bound on dim(Xc : 0) depending
on specific properties of the manifolds V and X.
Here is a result by A. Eremenko (see [Ere]), where V = C, X = CP m and the
maps we are concerned with are holomorphic ones.
The restriction map x 7→ x | 1, sending √ Xc → (CP m )1 , is an embedding,
provided 1 ⊂ C is ε-dense for ε < c−1 π/4. Furthermore,
dim(Xc : C) 6 2mC 2 /π ;
(d) More general (but less precise) results are available for harmonic maps,
where the elliptic estimates are controlled by the lower bound on Ricci curvature
of V and the upper bound on the sectional curvature of X as (apparently) follows
from the Yau gradient estimate (compare (b) in 3.4.2).
(e) Our embedding result states, in effect, that two distinct harmonic (line) maps
x and x 0 with bounded differentials cannot coincide on a sufficiently dense subset
6 in the manifold V where the maps are defined. Much more is known for holo-
morphic maps, where the first main theorem of the Nevanlinna theory provides a
bound on the density of 6 in terms of the growth of kDxk and kDx 0 k on V . This
leads to the following
GENERAL PROBLEM. Consider harmonic maps x from V to X or more general

maps satisfying some (linear or nonlinear) system of elliptic PDE. Take two non-
negative functions σ (v) and δ(v) en V and decide whether there exist two distinct
maps x and x 0 from our class, such that
max(kDx(v)k, kDx 0 (v)k) 6 σ 0 (v)
and
dist(x(v), x 0 (v)) 6 σ (v)
for all v ∈ V .
Here again, one expects the bound on a suitable density of the zero set of δ(v) in
terms of the asymptotic growth of σ (v) for v → δ. More generally, one wishes to
show, that if δ(v) is small on a rather dense set, then it is also small on a much larger
set, provided we have some bound on σ (v). For example, a holomorphic function
x with many zeros in a ρ-disk and with a bound on kx(v)k in the consecutive
2ρ-disk is much smaller on the ρ-disk than was suggested by the original bound
on kx(v)k. Another general phenomenon of this kind is the unique continuation
property for elliptic PDE but all this seems far away from a desirable solution of
the above problem.
3.4.4. Residual Dimension for Spaces of Holomorphic Maps

This refers to the dimension of the space of holomorphic maps x: V / 0i → X for
subgroups 0i ⊂ 0 of finite index. The above argument shows that the dimension of
the space of such maps x satisfyingkDxk 6 c is bounded by Acn |0/ 0i | for some
constant A = A(V , 0, X) and n = dimR V . In fact this remains valid for all our
harmonic-like maps while for holomorphic maps there is a better estimate due to
the following elementary (and well known)
3.4.5. PROPOSITION. Let X be a complex projective variety, W be a compact

connected complex manifold and let x0 : W → X be a holomorphic map. Then the
dimension of the connected component X0 of x0 in the space of holomorphic maps
W → X is bounded by the volume of the image of x0 and the maximal number
ν = ν(x0 ) of irreducible components of the fibers of x0 as follows
dim X0 6 Aν Vol2k x0 (W )
for k = dimC x0 (W ), ν = supx∈X card conn(x0−1 (x)), and some constant A =
A(X).
Proof. The dimension dim X0 is bounded by the dimension h of the space Hx of
holomorphic sections of the induced bundle x ∗ (T (X)) → W for a generic x ∈ X0 ,
as these Hx make up the tangent bundle of X0 on the nonsingular locus of X0 ,
which is known to be a complex variety in its own right. (Here we used smoothness
of X but this can be always achieved by embedding X into a smooth variety, e.g.
into a projective space.) Next we observed that T (X) can be embedded into a sum
of several very ample line bundles over X (this is true and obvious for all vector
bundles L over projective varieties) and the matter reduces to evaluation of the
dimension ` = dim H0 (x ∗ (L)). Such an L embeds X to some projective space
CP N and so we may think of x ∗ (L) as the restriction of the bundle O(1) to our W ,
now mapped to CP N by composing x: W → X and the embedding X → CP N .
Notice that the product ν(×) Vol ×(W ) is invariant under deformations of maps
and so all we need is to estimate ` for a map y0 : W → CP N in terms of ν =
ν(y0 ) Vol2k y0 (W ). We do this by induction on k as follows. Intersect y0 (W ) with a
generic hyperplane P and observe that our number ` = `k is bounded by `k−1 + `0
where `k−1 comes from P ∩ y0 (W ) and `0 is the dimension of the space of sections
of O(1) on W which vanish on P ∩ y0 (W ). This space easily identifies with the
space of sections of a trivial line bundle over y0 (W ) and so `0 = 1. Thus everything
reduces `0 where our variety consists of at most ν Vol y0 (W ) points counted with
multiplicity (for the usual in CP N , where the volume of each subvariety equals its
degree). Thus finally ` 6 ν Volzk y0 (W ) + k. 2
(Notice, this is sharp for the linear embeddings W = CP k → CP N .)

398 MISHA GROMOV
Remark. Probably the conclusion remains true for all complex (not necessarily
algebraic) X and, possibly, for more general harmonic (like) maps.
3.5. LOWER BOUNDS ON THE MEAN DIMENSION FOR SPACES OF

HOLOMORPHIC MAPS
If X is a compact Riemannian manifold and V is complete, then, typically, the

space of harmonic maps x: V → X with kDxk 6 const looks zero-dimensional
(probably, uncountable for many generic classes of metrics on V ), but I am not
aware of any published result of this kind. On the other hand, there are certain
remarkable exceptions, such as Káhler manifolds that sustain lots of holomorphic
maps and these are necessarily harmonic.
3.5.1. EXAMPLE: MAPS C → S 2 . These are just meromorphic functions x =

x(v), v ∈ C, which can be constructed in abundance with bounded spherical
derivatives as follows. Take a discrete subset 6 ⊂ C and consider meromorphic
functions ϕσ : C → C ∪ ∞ = S 2 = CP 1 of the form ϕσ (v) = cσ (v − σ )−k .
If the sum of these over all σ ∈ 6 converges, we get a meromorphic function
X: C → C ∪ ∞ = S 2 = CP 1 where one can easily control the differential dx.
For example, if 6 is separated, i.e. kσ1 − σ2 k > δ > 0 for all σ1 6= σ2 in 6 and
the coefficients cσ are bounded, then this sum obviously converges and gives us
an x: C → S 2 with supv∈C kdxk < ∞, provided k > 3. Moreover, by varying
cσ , one can easily make such an f with prescribed values on a sufficiently rare net
6 0 ⊂ Cn lying away from 6. This shows, that the space Xc of holomorphic (and
thus harmonic) maps x: C → S 2 with kdxk 6 c has
dim(Xc : C) = κc2 for some κ > 0.
Consequently, if a complex analytic manifold X contains a rational curve then
the space of holomorphic maps x: C → X with kdxk 6 c has positive mean
dimension for all c > 0.
Remark. By varying σi and/or rotating the sphere S 2 , one can easily make an
x: C → S 2 with kdxk 6 c and prescribed values on a given, sufficiently sparce
(depending on c > 0) net 6 0 ⊂ C. (See [Ere] for a finer construction of such
interpolating maps C → CP n .)
3.6. L2 - TECHNIQUE FOR MAPS V → CP N
Let V = (V , g) be a Hermitian manifold with locally bounded geometry and E →

V a strictly positive line bundle, i.e. with the curvature form w satisfying w > λg.
3.6.1. EMBEDDING THEOREM. There exists a holomorphic uniformly locally

bi-Lipschitz map x: V → CP N for some N = N(V , E).
Proof. First, for each point v ∈ V , we can construct n + 1 L2 -sections xj : E i

for some i and j = 0, . . . , n = dim V such that the map V → CP n defined by
these sections embeds some ball B(v, ρ) to CP n . Furthermore, by squaring xj , we
can make the sections L1 (see 2.1.3). Then, we take such sections at each point of
a sufficiently rare net 6 ⊂ V and by summing them up (compare 3.3.10), obtain
a map V → CP n that embed the ρ-neighbourhood Uρ (6) ⊂ V into CP n in a
S Finally we cover all of V by ρ-neighbourhoods of several
bi-Lipschitz manner.
such nets, V = ν Uρ (6ν ), ν = 1, . . . , N0 , and then the resulting map x: V →
CP nN0 +N0 −1 is clearly seen to be locally bi-Lipschitz. 2
Remark. If V is compact, the above amounts to the classical Kodaira theorem,

where one can, moreover, project V from CP N to CP 2n+1 and then further to
CP n if one is not concerned so much with embeddings. But if V is noncompact,
the image of V in CP N may be, a priori, dense and then there is no holomorphic
Lipschitz projection to CP N−1 . However, such projection can be obtained with the
uniform transversality theorem as will become clear later on.
3.6.2. Now, given a suitable holomorphic Lipschitz (i.e. with bounded differential)
map x0 : V → CP N we want to generate a larger space of such maps. To do this we
take the pull-back E → V of the O(1)-bundle over CP N and use bounded sections
of E i for this purpose. So we need E to be rather positive which is ensured by the
following condition generalizing the ‘locally bi-Lipschitz’ property.
Uniform nondegeneracy. Let x: V → CP N be a holomorphic Lipschitz map.

Since V has bounded geometry, we have a local coordinate system with ‘bounded
distortion’ at each point v ∈ V and so by looking at x and on all small balls in V
we obtain a precompact family of holomorphic maps from the unit ball B ⊂ Cn
to CP N , call them xv : B → CP N . We say that x is uniformly nondegenerate if
every map y: B → CP N belonging to the closure of the family {xv }, v ∈ V ,
(with the uniform topology) is finite to one. For example, if V is compact, then this
equivalent to x itself being finite to one.
Now, it is essentially standard that if x: V → CP N is uniformly nondegen-
erate holomorphic Lipchitz map, then then induced Hermitian structure in E =
x ∗ (O(1)) admits a small perturbation making the curvature of E strictly positive.
(Such a perturbation can be achieved, for example, along the stratification of the
locus where the differential Dx is noninjective.)
3.6.3. PROJECTIVE INTERPOLATION THEOREM. Suppose V admits a uni-

formly nondegenerate holomorphic Lipschitz map x0 to CP N . Then for every δ-
separated subset 6 ⊂ V there exists a holomorphic map x: V → CP N with
kDxk 6 const(1 + δ −1 ) taking given values at all points σ ∈ 6, where const =
const(V , x0 ).
Proof. The line bundle E = O(1) over CP N admits many (meromorphic) maps
into CP N different from the original projection. To see one, observe that each point
400 MISHA GROMOV
(vector) in E is given by a pair (`, ϕ) where ` ⊂ CN+1 is a line and ϕ: ` → C a

linear form.
Now, with a given vector z1 ∈ CN+1 , we associate the map pz1 : E → CP N
where the line `0 = pz1 (`) ∈ CP N is spanned by the vector z1 + ϕ −1 (1) ∈ CN+1 .
Notice, that this pz1 has poles, but it is regular in some (Zariski) neighbourhood of
the zero section CP N ⊂ L = x0∗ (E) of E to V , and observe that each holomorphic
section y1 : V → E with a sufficiently small sup-norm gives us a map of V to
CP N , that is the composed map pz1 ◦ y1 , denoted x1 : V → CP N . Furthermore, if
z1 6= 0, then the map pz1 is injective on each fiber of L near zero, and so we obtain
an embedding from the space of small sections V → E to the space of maps
V → CP N close to x0 . (Consequently, the mean dimension of the space of maps
x: V → CP N with kDxk 6 const is bounded from below by that for the space
of bounded sections V → E.) Then one can similarly deform x1 using some pz2
and y2 and so on. Thus the proof would be concluded if we had the interpolation
property in the bundle E.
We cannot guarantee that E itself has sufficiently many sections, but some
power E i is good for this purpose. To go from E to E i , we consider a selfmapping
ψ of CP N
√ given by polynomials of degree i such that ψ can be found with kDψk
about i (modeled on the standard map of CN /Z2N ) and it pulls back E to E i .
We compose ψ with our x0 : V → CP N and thus promote E to E i over V , as
(x0 ◦ ψ)∗ (E) = E i . Now we have as many sections as we need and the proof
trivially follows from 3.2.6. 2
COROLLARY TO THE PROOF (Compare 0.6.2). If V is acted upon by an
amenable group 0 with a projective algebraic quotient V / 0, then the space Xc
of holomorphic maps x: V → CP N with kDxk 6 c satisfies
dim(Xc : 0) > b0 cdimR V
for all sufficiently large c and some b0 > 0.
3.6.4. Projective Transversality Theorem

Let x0 : V → CP N be as above and consider a subset S in the space of r-jets
holomorphic maps V → CP N . We want to move x0 away from S, i.e., to have
the r-jet of the moved section to lie ε-far from S for some ε > 0. Again, we
cannot freely manipulate x0 itself, but we can work with xi = ψ ◦ x for the above
map ψ: CP N → CP N , where the above argument combines with the uniform
transversality theorem and leads to the following conclusion.
If the uniform dimension (see [GroPCMD ]) of S is strictly less than dimR V =
λ dimC V , then there exists a holomorphic R uniformly nondegenerate Lipschitz
map xi0 : V → CP N which is uniformly transversal to S.
COROLLARIES. (a) If N > dimC V , then one can make the xi0 miss a small ball in
CN N . This allows projections from CP N to CP N−1 and eventually to CP n . Thus
we obtain a holomorphic uniformly nondegenerate map V → CP n , n = dim V .
(b) If N > 2n − 1, we can produce uniform immersions V → CP N , i.e.

uniformly locally bi-Lipschitz maps.
3.6.5. Remarks and questions. (a) If V = Cn or if V admits a nonconstant

holomorphic map V → Cn with bounded differential, then there are lots of holo-
morphic maps x: V → CP n with kDxk 6 c for arbitrarily small c > 0. On the
other hand, for some V , every map x: V → CP n with sufficiently small kDxk
is necessarily constant. This is the case, for example, for infinite cyclic coverings
of compact manifolds as well as for more general V which are one-dimensional at
infinity in the sense of [GroPCMD ]. Can one classify manifolds with this properly?
Similarly, assuming V is acted upon by an amenable group 0, what is the mean
dimension of the space Xc of holomorphic maps V → CP N for small c < 0?
Now, in general, does the dimension dim(Xc : 0) depend on c, especially for
c → 0?
(b) What is the relation of dim(Xc : 0) and the corresponding residual dimen-
sion for residually finite groups 0? In particular, when can a holomorphic map
x: V → CP N with kDxk 6 c be approximated by 0i -periodic maps xi with
kDxi k 6 c1 where 0i ⊂ 0 is some sequence of subgroups of finite index and cj
is independent of i? Closely related to this is the Runge approximation problem
where we look for an approximate extension of holomorphic Lipschitz maps from
smaller domains in V to larger ones. Finally, one asks when holomorphic Lipschitz
maps to CP N extend from subvarieties W ⊂ V with bounded local geometry to
all of V .
(c) Foliations. Consider a manifold U (or a general locally compact space for
this matter) foliated into complete Hermitian manifolds V and let E → U be
a complex line bundle holomorphic along the leaves. For each point u ∈ U we
take the universal covering V eu of the leaf Vu ⊂ U passing through u, thought of
as the space of the homotopy classes of loops in Vu ⊂ U based at u so that u
canonically lifts to Veu and is denoted ũ ∈ V ev . Let H
eu be the space of holomorphic
L2 -sections of the bundle E eu → V eu induced from E and take a vector e in the (one-
dimensional) fiber Eu ⊂ E at u identified with the corresponding fiber Eũ of E eu . If
e e
the evaluation map Hu → Eũ = Eu is surjective, then there exists a unique section
h̃ = h̃e : Veu → E eu having h̃(u) = e and minimizing the L2 -norm kh̃kL2 . Now,
suppose E is positive along the leaves, where this positivity is uniform on U, and
also assume that all leaves have uniformly bounded local geometry. Then the above
surjectivity condition is satisfied for all e ∈ Eu and all u, if not for E itself, then,
at least, for some tensorial power E i of E. Thus every section ϕ: u 7→ e(u) ∈ Eui
of E i gives rise to a family of holomorphic L2 -sections h̃u =def h̃ϕ(u) : V eu → E eui ,
that is a section, call it e
ϕ , of E i lifted to the graph Ve of our foliation defined as the
e
space of pairs (u, ṽ), for u ∈ U and ṽ ∈ Vu . According to 3.3.12, each holomorphic
constituent h̃u of eϕ has a controlled decay on the leaf V eu with the decay estimate
independent of u. Moreover, the L2 -estimate (see (∗) in 3.3.2) implies that the
sections h̃u are L2 -continuous in u, provided ϕ is continuous. In fact, if u and u0 are
402 MISHA GROMOV
near by points in U and e ∈ Eui and e ∈ Eu0 0 are close vectors, then the leaves V eu and
eu0 are close on large balls B
V eu ⊂ Veu and B eu0 ⊂ V
eu0 . Then the holomorphic section
e
h̃e , on Bu0 can be moved to an almost holomorphic section h̃0 on Bu which lies close
to h̃e0 on B eu0 and has h̃0 (u) = e. This h̃0 can be made zero outside B eu by applying
an obvious cut-off argument and then we observe that the section h̃00 = 12 (h̃e + h̃0 )
is also almost holomorphic and thus can be turned holomorphic as we did before
using the L2 -estimate. The resulting holomorphic h̃• is L2 -close to h̃00 and may be
assumed to have h̃• (u) = e. On the other hand, if h̃e0 were far from h̃e , then h̃0 is far
from h̃e as well, and then their mean would have significantly smaller L2 -norm than
h̃e (as we could assume kh̃e0 kL2 6 kh̃e kL2 by interchanging u and u0 otherwise)
which contradicts to minimality of the norm kh̃e kL2 . Next, we square each h̃u thus
making it L1 and then push down the resulting section e e to a section ϕ 2 (v)
ϕ 2 from V
of E → U by integrating e
2i
ϕ (u, ṽ) with respect to u over the leaf Vv = Vu and
2
e
summing up over all ṽ in Vu over v, where we need ϕ to be bounded (as well as
continuous) on U. In particular, if U is compact, we obtain, by varying ϕ, lots of
continuous sections of E 2i → U holomorphic along the leaves and conclude that
U admits a continuous leafwise holomorphic and leafwise locally bi-Lipschitz map
to some CP N .
Notice, that such foliations exist, for example, on locally homogeneous spaces
U of the form K\G/ 0 where G is a semi-simple group without compact quotients,
K ⊂ G a (nonmaximal!) compact subgroup and 0 ⊂ G is a cocompact lattice.
(d) Singular spaces V . Probably, our results extend to singular spaces V with
an obvious extension of the idea of bounded local geometry. For example, one can
easily handle submanifolds V of a manifold W with bounded geometry, such that
Vol2n (V ∩ B) 6 const for all unit balls B in W . In general, one needs a suitable
version of ∂-technique where a natural idea is to embed V into a nonsingular mani-
fold. Alternatively, one may resolve the singularity of V and adjust the ∂-lemma to
sections constant (vanishing) on the pull-back of the singular locus. Alternatively,
one may try L2 -techniques on the Ĉech resolutions of the relevant sheaves.
(e) As we mentioned earlier, the space of harmonic maps between generic Rie-
mannian manifolds seems rather small but there are some exceptional cases besides
the Kähler manifolds. For example, one may look from this angle at harmonic maps
Rn → S N for all n > 2.ROne can also consider n-harmonic maps which locally
minimize the energy = kDxkp dv with p = n which bubble very much like
ordinary harmonic
R maps for n = 2. Here one should probably replace the uniform
metric for RkDxkp by ‘the energy metric’ and study maps x: V → X locally
minimizing kDxkp dv and having this integral uniformly bounded over the unit
balls in V . This is similar to bounding holomorphic maps x: V → X by their
‘local degrees’, i.e. by the volumes of their graphs within unit balls in V × X,
where one can use estimates from 4.1 or, alternatively, the (first main theorem of
the) Nevanlinna theory which, when it applies, gives better bounds on the mean
dimensions of these spaces than 4.1 (as was pointed out to me by Alex Eremenko).
Notice that in all these cases the spaces of maps with bounded local energy (or
degree) are not compact and one should compactify them by allowing suitable
‘singular maps’ best represented by certain subsets in V × X appearing as limits
of graphs of the maps in question.
(f) It is worthwhile to recall at this stage that the mean dimension of a space
of maps V → X appears as a limit of the ε-dimensions of spaces of maps of
relatively compact domains ⊂ V . A more general class of geometric problems
can be formulated for an arbitrary V , allowed to be noncomplete and/or to have a
boundary and for a relatively compact in the interior of V . Here we take some
space Xc of our (harmonic like) maps x: V → X with a bound like kDxk < c,
or a similar bound on the (local or global) energy of x. Then we restrict the maps
x ∈ Xc to and evaluate the ε-dimension Widimε of the resulting space Xc | of
maps → X with respect to some metric in this space, e.g. the uniform metric
or some energy metric. What we want to know is the asymptotic behavior of the
resulting Widimε (Xc |) for growing V and , where remains much smaller
than V . Here it is equally interesting to evaluate the minimal number Nε of the
ε-balls needed to cover Xc |, where the expected growth is roughly ε − Vol .
3.6.6. About Fusion

If dimC V = 1, then there are nonlinear techniques for producing holomorphic
maps V → X where X is an almost complex manifold (with possibly a noninte-
grable structure) which contains ‘sufficiently many’ rational curves. Here a given
holomorphic map x0 : V → x can be modified by ‘fusing’ it with rational curves
at the points σ of some discrete subset 6 ⊂ V . Recall that the analytic model
for ‘fusion’ of two curves c1 and c2 in X given by the equations f1 (x) = 0 and
f2 (x) = 0 is the curve C = Cε given by the equation f1 f2 = ε. This Cε for small
ε 6= 0 looks like the connected sum of C1 and C2 at their intersection points.
EXAMPLE. Let X = CP N with an almost complex structure tamed by the stan-

dard symplectic from w on CP N . One can easily construct, by fusing together
infinitely many rational curves, a holomorphic Lipschitz map x: C → CP N with
assigned values on a given separated subset 6 ⊂ C. Probably, there is a similar
interpolation result for all Riemann surfaces with bounded geometry. Also one
may try maps into more general spaces X, e.g., into rationally connected algebraic
manifolds X.
4. Spaces of Subvarieties
Let W be a Hermitian manifold and consider the space of complex analytic sub-
varieties M ⊂ W of given dimension n. All? possible topology in M comes from
the Hausdorff convergence on compact subsets in W . We shall use below a slightly
different topology induced on M from the space of currents on W . Namely, for
every collection of continuous forms w on W of degree 2n with compact support,
we set
404 MISHA GROMOV
Z Z

|M1 − M2 | = sup w− w .
w∈ M1 M2
Then Mi → M signifies that |M − Mi | → 0 for every finite collection . Notice

that the limits of Mi in this topology may acquire multiplicity. For example, the
graph Mλ ⊂ C2 of the function z → λz2 , converges, for λ → ∞, to the vertical
line in C2 with multiplicity two (while the Hausdorff limit gives us this line without
multiplicity. Our objective is the space Med consisting of the subvarieties M ⊂ W
of ‘local degree’ bounded by d. This means that
Vol2n (M ∩ B) 6 αW d
for all unit balls B ⊂ W and a suitable normalization constant αW > 0 which for
M = CN should be chosen equal the volume of the unit Euclidean 2n-ball. If W
has bounded geometry, then our study can be reduced to that in W = CN where
ed become more transparent.
the relevant properties of M ∈ M
4.1. NORMALIZATION AND CAUCHY INEQUALITY
A complex analytic subvariety M ⊂ CN can be locally represented as the graph of

multi-valued holomorphic map Cn → CN−n . Namely, for each point v ∈ M, there
exists a linear projection p: CN → Cn so that p is finite-to-one on M. Then a germ
of M at v becomes a ramified cover of a small ball B = B(p(v), ε) ⊂ Cn where it
is represented by the graph of a dv -valued map from B to CN−n . Such a map can
be viewed as a singled valued holomorphic map from B to the dv −, the symmetric
power of CN−n , say µ: B → Sdv CN−n and by Cauchy inequality we can bound
the differential of µ in a smaller ball, say in B 0 = B(p(v), ε/2) by something like
ε/2. (Notice that the variety Sdv CN−n is singular but it embeds into a smooth one
and so one can speak of norms of derivatives of maps into it.) Thus our objective
is a lower bound on ε in terms of the 2n-volume of M.
4.1.1. Controlled Normalization

We want to locate m-dimensional polydisks D m ⊂ CN (which are more suitable
for us than 2m-balls) for m = N − dim M, such that their intersections with M
are stable under ε-perturbations. Thus we say that D m is ε-transversal to M if its
boundary ∂D m is ε-far from M, dist(∂D m , M) > ε. Here every D m lies in some m-
dimensional affine subspace L ⊂ CN .√ Observe that this D m = (D(r))m contained
in the ball B(R) ⊂ C of radius R = nr and we call this R the radius of D m .
m
LEMMA. Let M be a complex analytic subvariety in CN of dimension n. Then for

every ρ > 0 there exists a polydisk D m in CN with the following properties.
(1) The center of D m is located at the origine of CN .
(2) The radius R of D m lies in the interval ρ/2 6 R 6 ρ.
(3) D m is ε-transversal to M, where

ε > constN ρ 2n+1 d −1 (ρ), (+)
for d = d(ρ) denoting the 2n-volume of the intersection of M with the ρ-ball
in CN around the origin.
Remark. One can show, a posteriori, that such D m exist for all R in the interval
1
2
ρ 6 R 6 23 ρ.
Proof. An obvious integral geometric (or, alternatively, transversality) argument

shows that almost all polycylinders D m centered at 0 ∈ CN have empty intersection
∂D m ∩ M. Here one may invoke the compactness property of analytic sets with
bounded volume and thus obtain a definite lower bound on dist(∂D m , M) for some
R. Next, as we want a quantitative result, we recall the relevant property of M
behind the compactness property which reads
Vol2n (M ∩ B(v, δ)) > αn δ 2n (∗)
for all v ∈ M, all balls B(v, δ) ⊂ CN at v and αn equal the volume of unit
Euclidean 2n-ball. Then we take some D m of radius R centered at O ∈ CN and
intersect it with the ε-neighbourhood Mε of M. We measure this intersection by
the minimal number N = N(D m , ε) of ε-balls needed to cover the part of this
intersection lying in the ‘band’ D m \ 12 D m = D m (ρ) − D m (ρ/2). If N 6 0.1ρ/ε,
then, clearly, there is an R between ρ/2 and ρ such that the boundary of D m (R)
misses Mε and (+) follows with this ε. Now, assume N > 0.1ρ/ε for all D m (R)
and bound the volume of M∩B(ρ) from below (∗) as follows. First, imagine we are
allowed the parallel translations of D m by distance ρ. Then we get about N(ρ/ε)2n
points in M with mutual distances > ε/2 and thus the volume of M covered by
these translated is bounded from below roughly by Nαn (ε/2)2n (ρ/ε)2n = Nαn ρ 2n .
It follows, N is bounded approximately by Vol(M ∩ B(ρ)/ρ 2n , hence 0.1ρ/ε 6
CN Vol(M ∩ B(ρ))/ρ 2n and (+) follows.
Finally, instead of translating D m (which we are not allowed to do as the center
of D m is fixed) we rotate it around some (m − 1)-plane L in Cm ⊃ D m . We
choose this L ⊂ Cm so that the significant part of the intersection Mε ∩ (D m \ 12 D m )
lies roughly ρ-far from L, i.e. the covering number for Mε ∩ (D m \ 12 D m \Lρ 0 ) is
> βN ρ/ε for some βN > 0 and ρ 0 > βN ρ. Granted such L, the rotation of D m
gives us essentially the same volume as the above translation. Finally, to see that
such L exists, we apply the same reasoning, but now we rotate L in Cm around
some (m − 2)-plane L0 ⊂ L. Then L0 is located with rotation of L0 around L00 and
so on down to a rotating line in C? 2
4.1.2. Local Representation of M by Multivalued Function

Consider the normal ε-tube around our D m that is D m × B0⊥ (ε) ⊂ CN=m+n where
B0⊥ (ε) is the n-dimensional ε-ball in CN normal to D m and let M0 denote the
406 MISHA GROMOV
intersection of M with this tube. Clearly, the projection of M0 to B0⊥ (ε) is a proper
map of multiplicity d0 = d0 (ε) 6 constn (Vol M ∩ B(ρ))/ε 2n. Thus M0 is rep-
resented by the graph of d0 -valued map over B0⊥ (ε), say ϕ0 : B0⊥ (ε) → Sd0 D m .
Then we consider such tubes centered at all points in M and cover M by a min-
imal number of these. Here we set ρ = 1 and denote by d the supremum of the
volumes of intersections of the unit balls in CN with M. Then we see with (∗)
that there is a covering of M by ε-tubes, where the number of such tubes meeting
each unit ball in CN is bounded by constN dε −2n 6 const0N d 2n+1 , where we use
(+) in the form ε > constN d −1 (and where we exercise the usual freedom with
the notation ‘const’N ). Finally we observe with (+) that d0 6 constN d 2n and so
dim Sd0 D m 6 constN md 2n 6 const0N d 2n . Thus the total number of ‘parameters
per unit volume’ defining M is bounded by constN d 4n+1 . This makes plausible
that the mean dimension of the space M ed of n-dimensional complex subvarieties
M ⊂ C with the local degrees bounded by d satisfies the inequality
N
ed : CN ) 6 constN d 4n+1 .
dim(M (?)
Actually, the natural conjecture (justified later on) reads
ed : CN ) 6 constN d n+1 ,
dim(M (??)
but we are not able to prove even the weaker inequality (?).
Here are two difficulties.
1. The above heuristic argument only applies to subvarieties close to a given
one and we lack a good localization theorem saying that ‘the local mean dimen-
sion equals the global one’. Thus we have to vary the tubes covering M which
unpleasantly enlarges the exponent 4n + 1 to something of order N 2 .
2. As we change a covering of M by ε-tubes, we change our representation of
M by a collection of maps (this already happens near a fixed M as we appeal to
Cauchy inequality). This introduces an ambiguity in our choice of a metric in M of
order d d (which probably could be greatly reduced) and this makes our exponent
(even in the local case) comparably large.
Remark. One can improve the covering argument of M by ε-tubes (using tubes
of variable size at different points in M) but I doubt you can bring the exponent
down to n + 1 this way. (The ‘difficult’ M’s are those having large intersections
with small balls, e.g., having conical singularities of degrees ≈ d.) On the other
hand, the above 1 and 2 are purely technical problems and should be eventually
resolved.
4.1.3. Parametrizations of Sd Cm
The symmetric powers of C are nonsingular. In fact, Sd C can be identified with C
in several ways. For example, given a symmetric configurationQof complex numbers
v1 , . . . , vd one can associate to them the polynomial p(z) = di=1 (z − vi ) and then
one uses the metric on Sd C corresponding to the sup-norm of functions on the disk
D(2) ⊂ C of radius 2. Another useful representation of Sd C is by means of the
symmetric functions,
X
d X
d X
d
s1 = mi , s2 = m2i , ..., sd = mdi ,
i=1 i=1 i=1
where the corresponding metric is the sup-norm in the (s1 , . . . , sd )-space. The two
norms are bi-Lipschitz equivalent, at least in the region corresponding to mi ’s with
|mi | 6 1 where the Lipschitz constant can be trivially bounded by something like
d d . Indeed, going from si to polynomials amounts to expressing the elementary
symmetric functions as polynomials in si . Conversely, one reconstructs si out of
p(z) by taking Cauchy integral of zi p 0 (z)/p(z) over the circle of radius 2 since
zi p 0 (z)/p(z) has simple poles at mi with residues mdi .
Next we observe that the natural map Sd Cm → (Sd C)m is finite-to-one, we
take, additionally, the composition of this projection with a generic linear map of
Cm . Then the resulting map Sd Cm → (Sd C)m × (Sd C)m = (Sd C)2m becomes
one-to-one.
4.1.4. Embedding of Md to a Power Space

We want to construct a sufficiently large set of m-disks in CN=m+n , so that each
M ⊂ M ed will be uniquely determined by intersections with these disks. (We
shall eventually disregard the disks which are not ε-transversal to M.) Recall, that
every m-disk in CN is of the form gD m (1) for the standard D(1) ⊂ C and some
isometry g: CN → CN . Thus we can mark the disks in our set by g’s. Here are our
requirements on these disks and g’s.
A. The set of g’s is ε0 -dense in the group IsomC CN for the standard metric,
where ε0 should be quite small, say ε0 6 const−1 (d+2)N 2
N (d + 2) for constN = N N .
B. If some m-disk D is in the set, then there is a δ-dense set of rotations of
this D = D m in the m-plane L spanned by D. That is g’s are δ-dense in the
subgroup (≈ U(m)) of unitary transformations of L fixing the center of D. Here
δ is independent of d, say δ = N −N . (Notice that the dependence of our constant
on N is a matter of convention as they could be absorbed by the definition of the
‘standard’ metric in Isom CN . Also observe, that the only role of this condition is
to take care of noninjectivity of the map Sd Cm → (Sd C)m .)
C. With every disk D in our family, there are ‘sufficiently many’ disks, say Di
obtained from D by parallel translation in the directions normal to D. Namely the
projection of this Di to the normal Cm is ε0 -dense, for the above ε0 in some ball of
radius 10 in Cm . (This is a purely technical condition. It is not truly needed but it
simplifies what follows.)
408 MISHA GROMOV
Clearly, there exists a system of disks with above properties, such that the num-
ber of these per unit volume in CN , i.e. meeting each unit ball (or cube) in CN does
not exceed constN ε0−K for
K = dim Isom CN + n = N(2N + 3) + n 6 3N 2 .
4.1.5. MAIN LEMMA. Let D be a collection of disks satisfying A, B, C and let
M1 and M2 be n-dimensional subvarieties in CN from the class M ed . Suppose, that
every disk D ∈ Dε-transversal to both M1 and M2 with the above ε, satisfies
D ∩ M1 = D ∩ M2 . Then M1 = M2 .
Proof. Since M1 ∪ M2 ⊂ M2d and our ε 0 is so small, we can cover CN by ε-
tubes g(D m × B(ε)) for D = D(1) ⊂ C, B(ε) ⊂ Cn and ε much larger than ε,
say ε = constN d −1 (see (+)) such that all disks gD m are in our collection and,
moreover, 2ε-transversal to both M1 and M2 . Thus, M1 and M2 are represented by a
collection of Sdg Cm -valued maps ϕg on the corresponding ε 0 -balls g(B(ε)) ⊂ CN ,
where dg are bounded by const0N d/(ε)2n 6 N 2N d 2n+1 . The intersection condition,
with B, says that these functions are equal on ε 0 -dense subsets in these balls. Now,
let δ denote the supremum of the distances between these functions over all our
balls. The argument as in Section 3 appealing to the Cauchy inequality makes
this distance λδ-small on concentric balls of radii, say 0.9ε with very small λ,
something of order 12 d −d . Thus δ-distance for one covering implies 12 d −d δ-distance
for another covering which then yield 12 δ-distance for the original covering by the
discussion in 3.1. It follows δ = 0 and the proof follows. 2
Remark. We did not try to be sharp in the above estimates but used notations
clarifying relative roles of n, N and d. Besides, there are little details to fill in, like
requiring covering by 0.8ε 0 -tubes (rather than the ε 0 -tubes), etc.
4.1.6. We want to interprete 4.1.5 as an embedding result and thus bound the mean
dimension of M ed . Denote by 1 = 1(D n ) the union of the cones of Si D m , i =
1, . . ., d0 joint at the vertex
_
d0
1= cone(Si D m ),
i=0
where d0 is the smallest integer > N 2N d 2n+1 , and let 1g = 1(gD m ). Now, for
every collections of disks gD m , g running over some subset g ∈ Isom CN , we
map Md to the Cartesian product ×g∈g 1g as follows. If gD m is 2ε-transversal to
M, then the g’s component of our map sends M to M ∩ gD m. If gD m is not ε-
transversal, we go to the joint vertex of the cones and we interplate between the
two maps in some standard way. Now Main Lemma shows that this map is an
embedding.
ed is bounded by
COROLLARY. The mean dimension of M
ed : CN ) 6 const(N, d),
dim(M (++)
const(N, d) 6 constN ε0−K dim 1.

Proof. All we need, is our collection of disks being Z2N -equivariant. Then we
found the mean dimension relative to Z2N that equals that for CN . 2
Remark. Our bound on K, ε and d0 = dim 1 are pretty awful. Better leave it
just as const(N, d).
4.1.7. The Proof of the Upper Bound in 0.6.4

The above proof of (++) is essentially local in nature and trivially generalizes to
subvarieties in all Hermitian manifolds with bounded local geometry. This gives us
the desired (horrible but effective) upper bound in 0.6.4. The lower bound will be
proven later on.
Remarks and open questions. (a) As we mentioned earlier, the constant in (++)
should be bounded by constN d 2n , where it will be interesting to explicitely com-
pute constN .
(b) The above argument can be, probably, extended to two-dimensional mini-
mal subvarieties in Riemannian manifolds and also to pseudo-holomorphic (one-
dimensional) subvarieties in almost complex manifolds W (where the easiest case
if of dimR W = 4 as we have at our disposal pseudo-holomorphic curves ε-
transversal to our M ⊂ W ). On the other hand, the situation seems more difficult
for higher dimensional minimal subvarieties. In fact, it seems unknown if the space
of n-dimensional minimal subvarieties of volume 6 d < ∞ in a compact Rie-
mannian manifold W has finite topological dimension. (On the other hand, generic
W ’s contain few minimal subvarieties and so, typically, their mean dimension
should be zero for infinite groups 0.)
(c) Clearly Med is empty for d 6 d0 = d0 (W ), where the critical d0 equals
1 for W = C . It is not hard to see that the mean dimension is continuous at
N
this critical value in the case of CN , dim(M ed : CN ) → 0 for d → 1, and,

probably, something similar holds true for all W . For example, if |K(W )| is small
and InjRad W > 1, then the critical d0 (W ) is close to one and the space M1+ε is
small f or small ε. In particular, if W is cocompactly acted by a discrete amenable
group 0, then dim(M1+ε : 0) → 0 for ε → 0 and |K(W )| → 0, as a simple
argument shows. (In fact, when d is close to 1, our M’s are uniformly nonsingular
and everything trivially reduces to linear PDE. Actually, this equally applies to
general minimal subvarieties with d 6 1 + ε, where the uniform nonsingularity
follows from Allard’s theorem. On the other hand, we do not know how to bound
the mean dimension of spaces of minimal varieties with d 1.)
410 MISHA GROMOV
4.2. ed AND RELATED QUESTIONS

RESIDUAL DIMENSION OF M
Let W be acted upon by a discrete group 0 with projective algebraic quotient W/ 0,

let 0i ⊂ 0 be a sequence of subgroups of finite index. The above discussion applies
to submanifolds in W/ 0i and shows, in particular, that the residual dimension
of Med is bounded by const(W, d, 0). We observe that 0i -invariant submanifolds
e
∈ Md descend to subvarieties in W/ 0i of volumes 6 const d|0/ 0i | and pose the
following
PROBLEM. Given a sequence of numbers δi , evaluate the dimensions of the

spaces Mδni (W/ 0i ) of n-dimensional subvarieties in W/ 0i of volume 6 δi .
Here we are interested in the asymptotic behavior of these dimensions for ‘in-
teresting’ sequences of subgroups 0i , where, specifically, we want to know the
answer for δi = |0/ 0i |α for a fixed α. We start with the following simple
OBSERVATION. Let W• be a compact N-dimensional manifold that admits a

holomorphic finite-to-one map ϕ: W• → CP N , such that the Kähler class of CP n
goes to a multiple of the Kähler class of W• say to λ[w(W• )] ∈ H 2 (W• ; R), then
dim Mδ 0 (CP N ) 6 dim Mδ (W• ) 6 dim Mδ 00 CP N
for δ 0 = λn (deg ϕ)δ and δ 00 = λn δ where deg ϕ denotes the toplogical degree of ϕ.
In fact, Vol ϕ −1 (M) = λ2n (deg ϕ) vol M for all M ⊂ CP n , which yields the
lower bound on dim Mδ (W ), while Vol ϕ(M) = λn Vol M for all M ⊂ W• which
gives us the upper bound.
Remarks. (a) This observation applies, strictly speaking, only to those Wi =

W/ 0i where 0i acts freely on W in order to have Wi nonsingular. But everything
(and obviously) equally works in the singular case.
(b) The above inequalities are most efficient for small deg ϕ but for our Wi =
W/ 0i we only guarantee maps ϕi : Wi → CP N with deg ϕi = const |0/ 0i | and
one cannot do better in most (?) cases, e.g., for the groups 0 satisfying Kazhdan’s
property T (see [GroMIKM ]). On the other hand, there are cases where deg ϕi 6
const independently of i, e.g., for coverings of an Abelian variety.
Now we recall the standard bounds for dim Mδn (CP N ).
4.2.1. LEMMA. The space Mδn1 (CP N ) of irreducible n-dimensional subvarieties

in CP N of degree δ satisfies
(δ + 1)(δ + 2) . . . (δ + n + 1)
−1
(n + 1)!

(δ + 1)(δ + 2) . . . (δ + n + 1)
6 dim Mδn (CP N ) 6 (N − n) −1
(n + 1)!
6 constN δ .
n+1
Proof. If N − n = 1 then M’s are given by homogeneous polynomials of degree

δ and so
(δ + 1) . . . (δ + n)
dim Mδn (CP N ) = .
n!
Next, if N − n > 2, we project M to N − n (n + 1)-planes in general position in
CP N and observe that M appears as an irreducible component of the intersection
of the pull-backs of these components. 2
EXAMPLE (Abelian varietes). Let 0 be a Lattice acting on CN with projective

algebraic quotient W• = CN / 0, e.g. 0 = i Z2N , i = 1, 2, . . .. Then there is our
map ϕ: W• → CP N with deg ϕ 6 const < ∞ independently of 0 with λ =
(Vol W• )−1/N (where the Kähler metric in CP N is normalized to have Vol CP N =
1). Then dim Mδn (W• ) is approximately (i.e. up to a multiplicative constant) equal
to
n(n+1)
(δλn )n+1 = δ n+1 (Vol W• )− N .
Thus, if we set d = δ/ Vol W• , we get

(n+1)(N−n)
dim Mδn (W• ) ∼ d n+1 (Vol W• ) N . (∗)
If N = n + 1, this becomes
dim Mdn (W• ) ∼ d n+1 Vol W•
and gives us the following bound on the residual dimension of the space M ed =
e
Md (C ) (of n-dimensional subvarieties M with Vol M ∩ B(1) 6 d for all unit
N
balls B(1)), resdim M ed 6 const d n+1 . This improves our earlier bound (with a poor
dependence on d) and suggests that the mean dimension of M ed must be asymptotic
to d n+1 . Here is a more general
CONJECTURE. Let W be a Hermitian manifold of bounded local geometry and

{Bi }i∈I be a collection of balls of radii ri 6 1, such that the concentric balls of
radii ri /2 cover W . Consider the space M of n-dimensional subvarieties M ⊂
W , such that Vol(M ∩ Bi ) 6 di rin for all i and given di > 0. Then dim M 6
P
const i∈I din+1 , where the constant depends only on N = dim W and the implied
bound on the local geometry of W .
The above conjecture truly makes sense only for compact W , where in general
one should use a suitable ‘dimension per unit volume’ in W . For example, if W
is cocompactly acted upon by an amenable group 0 and the system {Bi } is 0-
P then the mean dimension dim(M : 0) should be bounded by
invariant,
const i din+1 for i running over a fundamental domain J ⊂ I , i.e. a subset such
that 0J = I .
412 MISHA GROMOV
Remark. For certain manifolds W , e.g., for W = RN , it is interesting to look at

M’s defined with systems of balls Bi where ri are unbounded, say for concentric
balls of radii i → ∞ in CN . Then one may try to evaluate some ‘asymptotic
dimension’ of M in the spirit of the Nevanlinna theory. For example, let Mϕ denote
the space of n-dimensional subvarieties M in CN , such that vol(M ∩B(R)) 6 ϕ(R)
for a given function ϕ(R) and all concentric R-balls B(R) ⊂ CN around the origin.
Denote by Mϕ (R) the space of subvarieties in B(R) of the form M ∩ B(R) for all
M ∈ Mϕ . Then one may ask what is the asymptotic behaviour of dimε Mϕ (R)
for R → ∞ (and eventually for ε → 0) with respect to the Hausdorff metric in
Mϕ (R). A particularly interesting case is ϕ(R) = CR p for some p > n. (If p = n,
M is necessarily algebraic and so dimε Mϕ (R) is uniformly bounded.)
Now, let us look at the above asymptotic relation (∗) for codim M > 2, i.e. for
N − n > 2. Here the exponent (n + 1)(N − n)/2 is strictly greater than 1, and so
(∗) yields no bound at all on the residual (as well as on mean) dimension of M ed .
However, this does not contradict (++) from 4.1.7 but rather shows that majority
of subvarieties M ⊂ W are highly nonuniformly distributed in W for N − n > 2
and so (++) does not apply. This suggests the following
ALGEBRAIC QUESTIONS. Consider the space Mδn (CP N ) of algebraic subvari-

eties in CP N of dimension n and degree δ. How many irreducible components of
Mδn (CP N ) lie in the interval [δ α1 , δ α2 ] for given 0 < α1 < α2 6 n + 1? Here we
are most interested in the asymptotic behaviour of this number for δ → ∞, where
a good answer is plausible for large α1 , e.g., α1 > n.
To get some perspective look at the space M ⊂ Mδn (CP N ) of complete inter-
sections M of hypersurfaces of degrees δ1 > δ2 > · · · > δN−n . Its dimension is
easy to evaluate by looking at the normal bundle of M or by rescaling CP N by δ1
(which makes the volume of hypersurfaces of degree δ1 equal that of the rescaled
CP N and then applying (++) to the rescaled picture). Thus one easily shows that
dim M ∼ δδ1n and so each M ∈ M is contained in a hypersurface of degree, δN−n
which is roughly bounded by
1
n + 1 1 N−n−1
δ /d n for d = dim M (since δ = δ1 δ2 . . . δN−n ).
n
This suggests that for every irreducible variety M ⊂ Mρn (CP N ) of (large) dimen-
sion D, each M ∈ M is contained in a hypersurface of degree 6 δ 0 , where δ 0 can
be (reasonably) evaluated in terms of D and δ. For example, if D > εδ n+1 , then
one expects δ 0 6 ε 0 = ε 0 (ε, N), and if D > δ n+1−α for a small α > 0, then δ 0 < d
for δ > δ0 = δ0 (α, N). (Notice that holomorphic maps CP n → CP N with images
of degree δ make a variety of dimension about δ n whose generic members do not,
apparently, lie in hypersurfaces of degrees < δ and so ‘small’ should be at least
‘smaller than one’.)
There is another idea also expressing nonuniform distribution of subvarieties of
codimension > 2 in CP N . For example, one may seek a nontrivial upper bound
on the dimension of the subspace Nδn ⊂ (CP N )k consisting of those k-tuples of

points which lie on some subvariety M ⊂ CP N of dimension n and degree δ. For
example, if k > εδ n+1 , then, probably, codim Nδn > 0 for all δ > δ0 = δ0 (ε, N). In
fact our inequality (++) suggests, that no configuration of points (x1 , . . . , xk ) ∈
Nδn can be uniformly dense in CP N , i.e. the ρ-neighbourhood of {x1 , . . . , xk } ⊂
CP N must have small measure for ρ not much exceeding (εδ n+1 )−N and δ >
δ0 (ε, N).
4.3. CONSTRUCTION OF SUBVARIETIES IN W
Let W admit a positive line bundle E of locally bounded geometry. Then W admits
a holomorphic uniformly nondegenerate Lipschitz map x to CP N , N = dim W .
The pull-backs of subvarieties in CP N are, clearly, in our class M ed→∞ and by
e
varying x one sees that Md has positive mean dimension. Actually, by a direct ap-
plication of the uniform transversality theorem one obtains bounded holomorphic
sections x: W → E i which are uniformly transversal to the zero section in the
obvious sense. The zero set x −1 (0) ⊂ W of such an X is a manifold with bounded
local geometry of dimension equal dim W − 1 and so one obtains by induction
such submanifolds of all codimensions. This combines with an obvious scaling ar-
gument and shows, in particular, that in the presence of cocompact amenable action
the mean dimension of the space M ed of n-dimensional submanifolds M ⊂ W with
the bound Vol B ∩ M 6 d for all unit balls B ⊂ W satisfies
ed : 0) > const d n+1
dim(M
for all n 6 N, some const = const(W, 0) > 0 and all sufficiently large d.
Remarks and final questions. Since every complex subvariety is minimal, one
sees with the above theorem, for example, that the space Md of 2m-dimensional
minimal subvarieties M ⊂ RN with the volume bound by Vol M ∩ B 6 d has
dim(Md : RN ) > 0 for all N > 2m + 2 and d > Vol B 2m . But it is unclear if this
dimension is positive for minimal surfaces in R3 (where one can use the Weirstrass
representation to generate minimal surfaces).
Another situation where one may expect positive mean dimension is that of
pseudo-holomorphic subvarieties M ⊂ W with dimR M = 2, but here one needs
a different technique for producing sufficiently many of them in suitable almost
complex manifolds W .
Finally, we mention special Lagrangian submanifolds and related classes of
complex submanifolds, e.g., M ⊂ CN isotopic relative to a given (symmetric or
anti-symmetric) bi-linear form on CN . Unfortunately, the lack of examples pre-
cludes us from asking meaningful questions.
414 MISHA GROMOV
References
[Con] Connes, A.: Noncommutative Geometry, Academic Press, San Diego, 1994.
[Delz] Delzant, T.: Sur l’anneau d’un groupe hyperbolique, C.R. Acad. Sci. Paris Sér. I
Math. 324 (1997), 381–384.
[Dod-Mat] Dodziuk, J. and Mathai, V.: Approximating L2 invariants of amenable covering
spaces: A combinatorial approach, J. Funct. Anal. 154(2) (1998), 359–378.
[Dra] Dranishnikov, A.: Homological dimension theory, Russian Math. Surveys 43 (1988),
11–63.
[Ere] Eremenko, A.: Normal holomorphic curves from parabolic regions to projective
spaces, Preprint, 1998.
[Gal-Mey] Gallot, S. and Meyer, D.: Opérateur de courbure et laplacien des formes dif-
férentielles d’une variété riemannienne, J. Math. Pures Appl. (9) 54(3) (1975),
259–284.
[Gri-Ha] Griffiths, P. and Harris, J.: Principles of Algebraic Geometry, Wiley, New York, 1978.
[GroAI ] Gromov, M.: Asymptotic invariants of infinite groups, In: Niblo and Roller (eds),
Geometric Group Theory, Vol. 2, Proc. Symp. Sussex Univ., Brighton, July 14–
19, 1991, London Math. Soc. Lecture Notes Ser. 182, Cambridge Univ. Press,
Cambridge, 1993, pp. 1–295.
[GroCC ] Gromov, M.: Carnot–Carathéodory spaces seen from within sub-Riemannian
geometry, In: A. Bellaiche (ed.), Proc. journées nonholonomes; géométrie sous-
riemannienne, théorie du contrôle, robotique, Paris, June 30–July 1, 1992, Progr.
Math. 144, Birkhäuser, Basel.
[GroCDB ] Gromov, M.: Curvature, diameter and Betti numbers, Comm. Math. Helv. 56 (1981),
179–195.
[GroESAV ] Gromov, M.: Endomorphisms of symbolic algebraic varieties, J. Europ. Math. Soc.
(1999), 109–197.
[GroFRM ] Gromov, M.: Filling Riemannian manifolds, J. Differential Geom. 18 (1983), 1–147.
[GroMIKM ] Gromov, M.: Metric Invariants of Kähler manifolds, In: Caddeo and Tricerri (eds),
Proc. Workshops on J. Differential Geom. and Topology, Alghero, Italy, 20–26 June
1992, World Scientific, Singapore, 1993, pp. 90–117.
[GroNLS ] Gromov, M. and Piatetski-Shapiro, I.: Non-arithmetic groups in Lobatchevski
spaces, Publ. Math. IHÉS 66 (1988), 27–45.
[GroPCMD ] Gromov, M.: Positive curvature, macroscopic dimension, spectral gaps and higher
signatures, In: Gindikin and Simon (eds), Functional Analysis on the Eve of the
21st Century, Volume II. In honor of the eightieth birthday of I. M. Gelfand. Proc.
Conf. Rutgers Univ., New Brunswick, NJ, U.S.A., Oct. 24–27, 1993. Progr. Math. 132,
Birkhäuser, Basel, 1996, pp. 1–213.
[GroPDR ] Gromov, M.: Partial Differential Relations, Ergeb. der Math. 3. Folge, Bd. 9,
Springer-Verlag, New York, 1986.
[G-L-P] Gromov, M., LaFontaine, J. and Pansu, P.: Metric Structures for Riemannian and
Non-Riemannian Spaces, Based on Structures métriques des variétés riemanniennes,
edited by J. LaFontaine and P. Pansu, English translation by Sean Michael Bates,
Progr. Math. 152, Birkhäuser, Basel, 1999.
[Hed] Hedlund, G.: Endomorphisms and automorphisms of the shift dynamical system,
Math. Systems Theory 3 (1969), 320–375.
[Lin] Lindenstrauss, E.: Mean dimension, small entropy factors and an embedding theo-
rem, Preprint, 1998.
[Lin-Wei] Lindenstrauss, E. and Weiss, B.: Mean topological dimension, to appear in Israel J.
Math.
[Lück] Lück, W.: L2 -invariants of regular coverings of compact manifolds and CW -
complexes, Preprint.
[Orn-Weis] Ornstein, D. and Weiss, B.: Entropy and isomorphism theorems for actions of
amenable groups, J. Anal. Math. 48 (1987), 1–141.
[Nap] Napier, T.: Convexity properties of coverings of smooth projective varieties, Math.
Ann. 286(1–3) (1990), 433–479.
[Pass] Passman, D. S.: The Algebraic Structure of Group Rings (Reprint with corrections of
the orig. edn publ. 1977 by Wiley, New York), Krieger Publ., Florida.
[Schm] Schmidt, K.: Dynamical Systems of Algebraic Origin, Progr. Math., Birkhäuser,
Basel, 1995.
[Tian] Tian, C.: On a set of polarized Kähler metrics on algebraic manifolds, J. Differential
Geom. 32 (1990), 99–130.
[Ve-Go] Vershik, A. and Gordon, E.: Groups which locally embed into the class of finite
groups, Algebra Anal. 9(1) (1997), 71–97 (in Russian).
Contents of Volume 2
Volume 2 No. 1 1999
LI-YENG SUNG / Square Integrability and Uniqueness of the Solutions

of the Kadomtsev–Petviashvili-I Equation 1–24
WERNER KIRSCH and VLADIMIR KOTLYAROV / Soliton Asymp-
totics of Solutions of the Sine-Gordon Equation 25–51
NAKAO HAYASHI and PAVEL I. NAUMKIN / On the Davey–Stewartson
and Ishimori Systems 53–81
P. BUSCH / Stochastic Isometries in Quantum Mechanics 83–106
Instructions for Authors 107–112
Volume 2 No. 2 1999
L. BOUTET DE MONVEL / Complex Star Algebras 113–139

S. YU. DOBROKHOTOV and A. I. SHAFAREVICH / “Momentum”
Tunneling between Tori and the Splitting of Eigenvalues of the
Laplace–Beltrami Operator on Liouville Surfaces 141–177
PASQUALE GIOVINE / Nonclassical Thermomechanics of Granular
Materials 179–196
DANIEL H. LENZ / Random Operators and Crossed Products 197–220
Volume 2 No. 3 1999
MICHAEL DEMUTH and KALYAN B. SINHA / Schrödinger Opera-

tors with Empty Singularly Continuous Spectra 223–244
FRANZ MERKL / An Asymptotic Expansion for Bloch Functions on
Riemann Surfaces of Infinite Genus and Almost Periodicity of
the Kadomcev–Petviashvilli Flow 245–278
CONTENTS OF VOLUME 2
PETER STOLLMANN / Lifshitz Asymptotics via Linear Coupling of

Disorder 279–289
LECH ZIELINSKI / Sharp Spectral Asymptotics and Weyl Formula for
Elliptic Operators with Non-smooth Coefficients 291–321
Volume 2 No. 4 1999
MISHA GROMOV / Topological Invariants of Dynamical Systems and

Spaces of Holomorphic Maps: I 323–415
Volume Contents 417–418
Instructions for Authors 419–424

V. - A. - Marchenko - A. - Boutet - de - Monvel - H. - McKean - ( (BookFi) PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

V. - A. - Marchenko - A. - Boutet - de - Monvel - H. - McKean - ( (BookFi) PDF

Încărcat de

Drepturi de autor:

Formate disponibile

Mathematical Physics, Analysis and Geometry 2: 1–24, 1999.

Square Integrability and Uniqueness of the

(Received: 20 February 1998; in final form: 27 November 1998)

Mathematics Subject Classification (1991): 35Q53.

(qt − 6qqx + qxxx )x = 3qyy , (1.1a)

(g) The operators P± are defined by

They always act on functions in the k variable in this paper.

2. The Inverse Spectral Method

iφy − φxx + q0 φ = 0, (2.1)

and introduce the Jost function µ(x, y, k) by the relation

φ(x, y, k) = µ(x, y, k)e−i(xk−yk ) .

The equation for µ is

iµy − µxx + 2ikµx + q0 µ = 0. (2.3)

Taking the Fourier transform in the x variable, we can write (2.3) as

where δ is the Dirac function, and E± are defined by

The left scattering data L± (k, l) (k, l ∈ R) are defined by

µ̃l (ξ, y, k) = 2π δ(ξ ) +

where the integrals exist as iterated integrals.

3. The Equivalence of the Left and Right Solutions

LEMMA 3.1. The operators I − L± and I − R± are bounded and invertible on

Let (L± )t and (R± )t be the transposes of L± and R± respectively, i.e.,

We have the following analog of Lemma 3.1.

γ l (x, y, k) = e−i(xk−yk ) ηl (x, y, k),

Combining (2.16), (3.15) and (3.16) we have

The first limit in (3.13) follows from (3.14) and (3.17). 2

LEMMA 3.4. Let f (k) ∈ S(R), then

LEMMA 3.5. The following identity holds as operators on L2 (R):

Proof. From (3.9) we have (I − R− )−1 (I − L− )φ l = (I − R+ )−1 (I − L+ )φ l . Let

G± = [I − (L± )t ][I − (R± )t ]−1 g. (3.22)

Since G± ∈ L1 (R) by Lemma 3.2 and φ l (x, y, k) is bounded and continuous in

It follows from (3.23) that

where [E(x,y,t )f ](k) = ei(xk−yk +4t k ) f (k).

Combining (3.20), (3.25) and (3.26) we obtain the following lemma.

LEMMA 3.6. The functions λ(x, y, t, k) and ρ(x, y, t, k) are related by

= E(x,y,t )(I − R− )−1 (I − L− )E−1

The proposition follows from (3.27)–(3.29). 2

4. The Integrability of λ(x, y, t, k) and ρ(x, y, t, k) in the (x, k) Variables

The lemma follows from (4.2) and the smallness of Cq0 . 2

sP+ (e−isk f (k)) = −iP+ (e−isk f 0 (k)) + i[P+ (e−isk f (k))]0 .

For m = 0 or 0 6 s < 1, this follows immediately from the properties of the

It follows from (4.4) and (4.5) that

Hence we also have kg1 kLp . kf kWpm+1 . 2

Moreover, if B ⊂ R2 is bounded and ε > 0, then

LEMMA 4.4. Let V (x, y, t, k) be either

Then the following estimate holds:

lim kV (x, y, t, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B, (4.18)

r = {(x, k) ∈ R2 : |x| > r}. (4.19)

where G is defined in (4.13).

LEMMA 4.5. Let H (x, y, t, k) be as in Lemma 4.3. Then

lim k(L(y,t )H )(x, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B. (4.23)

uniformly for ε < |ξ | < ε −1 and (y, t) ∈ B as r → ∞.

LEMMA 4.6. The following estimate holds:

lim kλ(x, y, t, k) − 1 − (L(y,t )1)(x, k)kL2 (r ) = 0

χ(y,t ) = (1 + |y|)L2(y,t )1 + L(y,t )χ(y,t ). (4.33)

ω(y,t ) = L2(y,t )1 + L(y,t )ω(y,t ). (4.34)

LEMMA 4.7. The following estimates hold:

where Pj is a polynomial. Moreover, if B is a bounded subset of R2 and r is

Proof. Differentiating (2.10) we find

λx = h− + h+ + L(y,t )λx , (4.37)

We can write h− (x, y, t, k) as

We can estimate h1 (x, y, t, k) by Lemma 4.4, h2 (x, y, t, k) by Lemma 4.5 (with

The following proposition is obtained by using Lemmas 4.1–4.5, mathematical

5. The Square Integrability of q(x, y, t) in the (x, y) Variables

lim kV (x, y, t, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B, (4.18)

r = {(x, k) ∈ R2 : |x| > r}. (4.19)

lim k(L(y,t )H )(x, k)kL2 (r ) = 0 uniformly for (y, t) ∈ B. (4.23)

lim kλ(x, y, t, k) − 1 − (L(y,t )1)(x, k)kL2 (r ) = 0

where Pj is a polynomial. Moreover, if B is a bounded subset of R2 and r is